2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "compiler/glsl/ir.h"
26 #include "brw_fs_surface_builder.h"
28 #include "brw_program.h"
31 using namespace brw::surface_access
;
34 fs_visitor::emit_nir_code()
36 /* emit the arrays used for inputs and outputs - load/store intrinsics will
37 * be converted to reads/writes of these arrays
42 nir_emit_system_values();
44 /* get the main function and emit it */
45 nir_foreach_function(function
, nir
) {
46 assert(strcmp(function
->name
, "main") == 0);
47 assert(function
->impl
);
48 nir_emit_impl(function
->impl
);
53 fs_visitor::nir_setup_inputs()
55 if (stage
!= MESA_SHADER_FRAGMENT
)
58 nir_inputs
= bld
.vgrf(BRW_REGISTER_TYPE_F
, nir
->num_inputs
);
60 nir_foreach_variable(var
, &nir
->inputs
) {
61 fs_reg input
= offset(nir_inputs
, bld
, var
->data
.driver_location
);
64 if (var
->data
.location
== VARYING_SLOT_POS
) {
65 reg
= *emit_fragcoord_interpolation(var
->data
.pixel_center_integer
,
66 var
->data
.origin_upper_left
);
67 emit_percomp(bld
, fs_inst(BRW_OPCODE_MOV
, bld
.dispatch_width(),
69 } else if (var
->data
.location
== VARYING_SLOT_LAYER
) {
70 struct brw_reg reg
= suboffset(interp_reg(VARYING_SLOT_LAYER
, 1), 3);
71 reg
.type
= BRW_REGISTER_TYPE_D
;
72 bld
.emit(FS_OPCODE_CINTERP
, retype(input
, BRW_REGISTER_TYPE_D
), reg
);
73 } else if (var
->data
.location
== VARYING_SLOT_VIEWPORT
) {
74 struct brw_reg reg
= suboffset(interp_reg(VARYING_SLOT_VIEWPORT
, 2), 3);
75 reg
.type
= BRW_REGISTER_TYPE_D
;
76 bld
.emit(FS_OPCODE_CINTERP
, retype(input
, BRW_REGISTER_TYPE_D
), reg
);
78 int location
= var
->data
.location
;
79 emit_general_interpolation(&input
, var
->name
, var
->type
,
80 (glsl_interp_qualifier
) var
->data
.interpolation
,
81 &location
, var
->data
.centroid
,
88 fs_visitor::nir_setup_single_output_varying(fs_reg
*reg
,
89 const glsl_type
*type
,
92 if (type
->is_array() || type
->is_matrix()) {
93 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
94 const unsigned length
= glsl_get_length(type
);
96 for (unsigned i
= 0; i
< length
; i
++) {
97 nir_setup_single_output_varying(reg
, elem_type
, location
);
99 } else if (type
->is_record()) {
100 for (unsigned i
= 0; i
< type
->length
; i
++) {
101 const struct glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
102 nir_setup_single_output_varying(reg
, field_type
, location
);
105 assert(type
->is_scalar() || type
->is_vector());
106 this->outputs
[*location
] = *reg
;
107 this->output_components
[*location
] = type
->vector_elements
;
108 *reg
= offset(*reg
, bld
, 4);
114 fs_visitor::nir_setup_outputs()
116 if (stage
== MESA_SHADER_TESS_CTRL
)
119 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
121 nir_outputs
= bld
.vgrf(BRW_REGISTER_TYPE_F
, nir
->num_outputs
);
123 nir_foreach_variable(var
, &nir
->outputs
) {
124 fs_reg reg
= offset(nir_outputs
, bld
, var
->data
.driver_location
);
127 case MESA_SHADER_VERTEX
:
128 case MESA_SHADER_TESS_EVAL
:
129 case MESA_SHADER_GEOMETRY
: {
130 unsigned location
= var
->data
.location
;
131 nir_setup_single_output_varying(®
, var
->type
, &location
);
134 case MESA_SHADER_FRAGMENT
:
135 if (key
->force_dual_color_blend
&&
136 var
->data
.location
== FRAG_RESULT_DATA1
) {
137 this->dual_src_output
= reg
;
138 this->do_dual_src
= true;
139 } else if (var
->data
.index
> 0) {
140 assert(var
->data
.location
== FRAG_RESULT_DATA0
);
141 assert(var
->data
.index
== 1);
142 this->dual_src_output
= reg
;
143 this->do_dual_src
= true;
144 } else if (var
->data
.location
== FRAG_RESULT_COLOR
) {
145 /* Writing gl_FragColor outputs to all color regions. */
146 for (unsigned int i
= 0; i
< MAX2(key
->nr_color_regions
, 1); i
++) {
147 this->outputs
[i
] = reg
;
148 this->output_components
[i
] = 4;
150 } else if (var
->data
.location
== FRAG_RESULT_DEPTH
) {
151 this->frag_depth
= reg
;
152 } else if (var
->data
.location
== FRAG_RESULT_STENCIL
) {
153 this->frag_stencil
= reg
;
154 } else if (var
->data
.location
== FRAG_RESULT_SAMPLE_MASK
) {
155 this->sample_mask
= reg
;
157 int vector_elements
= var
->type
->without_array()->vector_elements
;
159 /* gl_FragData or a user-defined FS output */
160 assert(var
->data
.location
>= FRAG_RESULT_DATA0
&&
161 var
->data
.location
< FRAG_RESULT_DATA0
+BRW_MAX_DRAW_BUFFERS
);
163 /* General color output. */
164 for (unsigned int i
= 0; i
< MAX2(1, var
->type
->length
); i
++) {
165 int output
= var
->data
.location
- FRAG_RESULT_DATA0
+ i
;
166 this->outputs
[output
] = offset(reg
, bld
, vector_elements
* i
);
167 this->output_components
[output
] = vector_elements
;
172 unreachable("unhandled shader stage");
178 fs_visitor::nir_setup_uniforms()
180 if (dispatch_width
!= 8)
183 uniforms
= nir
->num_uniforms
/ 4;
187 emit_system_values_block(nir_block
*block
, fs_visitor
*v
)
191 nir_foreach_instr(instr
, block
) {
192 if (instr
->type
!= nir_instr_type_intrinsic
)
195 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
196 switch (intrin
->intrinsic
) {
197 case nir_intrinsic_load_vertex_id
:
198 unreachable("should be lowered by lower_vertex_id().");
200 case nir_intrinsic_load_vertex_id_zero_base
:
201 assert(v
->stage
== MESA_SHADER_VERTEX
);
202 reg
= &v
->nir_system_values
[SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
];
203 if (reg
->file
== BAD_FILE
)
204 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
);
207 case nir_intrinsic_load_base_vertex
:
208 assert(v
->stage
== MESA_SHADER_VERTEX
);
209 reg
= &v
->nir_system_values
[SYSTEM_VALUE_BASE_VERTEX
];
210 if (reg
->file
== BAD_FILE
)
211 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_BASE_VERTEX
);
214 case nir_intrinsic_load_instance_id
:
215 assert(v
->stage
== MESA_SHADER_VERTEX
);
216 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INSTANCE_ID
];
217 if (reg
->file
== BAD_FILE
)
218 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_INSTANCE_ID
);
221 case nir_intrinsic_load_base_instance
:
222 assert(v
->stage
== MESA_SHADER_VERTEX
);
223 reg
= &v
->nir_system_values
[SYSTEM_VALUE_BASE_INSTANCE
];
224 if (reg
->file
== BAD_FILE
)
225 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_BASE_INSTANCE
);
228 case nir_intrinsic_load_draw_id
:
229 assert(v
->stage
== MESA_SHADER_VERTEX
);
230 reg
= &v
->nir_system_values
[SYSTEM_VALUE_DRAW_ID
];
231 if (reg
->file
== BAD_FILE
)
232 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_DRAW_ID
);
235 case nir_intrinsic_load_invocation_id
:
236 if (v
->stage
== MESA_SHADER_TESS_CTRL
)
238 assert(v
->stage
== MESA_SHADER_GEOMETRY
);
239 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
240 if (reg
->file
== BAD_FILE
) {
241 const fs_builder abld
= v
->bld
.annotate("gl_InvocationID", NULL
);
242 fs_reg
g1(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
243 fs_reg iid
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
244 abld
.SHR(iid
, g1
, brw_imm_ud(27u));
249 case nir_intrinsic_load_sample_pos
:
250 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
251 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
252 if (reg
->file
== BAD_FILE
)
253 *reg
= *v
->emit_samplepos_setup();
256 case nir_intrinsic_load_sample_id
:
257 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
258 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
259 if (reg
->file
== BAD_FILE
)
260 *reg
= *v
->emit_sampleid_setup();
263 case nir_intrinsic_load_sample_mask_in
:
264 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
265 assert(v
->devinfo
->gen
>= 7);
266 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_MASK_IN
];
267 if (reg
->file
== BAD_FILE
)
268 *reg
= *v
->emit_samplemaskin_setup();
271 case nir_intrinsic_load_local_invocation_id
:
272 assert(v
->stage
== MESA_SHADER_COMPUTE
);
273 reg
= &v
->nir_system_values
[SYSTEM_VALUE_LOCAL_INVOCATION_ID
];
274 if (reg
->file
== BAD_FILE
)
275 *reg
= *v
->emit_cs_local_invocation_id_setup();
278 case nir_intrinsic_load_work_group_id
:
279 assert(v
->stage
== MESA_SHADER_COMPUTE
);
280 reg
= &v
->nir_system_values
[SYSTEM_VALUE_WORK_GROUP_ID
];
281 if (reg
->file
== BAD_FILE
)
282 *reg
= *v
->emit_cs_work_group_id_setup();
285 case nir_intrinsic_load_helper_invocation
:
286 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
287 reg
= &v
->nir_system_values
[SYSTEM_VALUE_HELPER_INVOCATION
];
288 if (reg
->file
== BAD_FILE
) {
289 const fs_builder abld
=
290 v
->bld
.annotate("gl_HelperInvocation", NULL
);
292 /* On Gen6+ (gl_HelperInvocation is only exposed on Gen7+) the
293 * pixel mask is in g1.7 of the thread payload.
295 * We move the per-channel pixel enable bit to the low bit of each
296 * channel by shifting the byte containing the pixel mask by the
297 * vector immediate 0x76543210UV.
299 * The region of <1,8,0> reads only 1 byte (the pixel masks for
300 * subspans 0 and 1) in SIMD8 and an additional byte (the pixel
301 * masks for 2 and 3) in SIMD16.
303 fs_reg shifted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
305 stride(byte_offset(retype(brw_vec1_grf(1, 0),
306 BRW_REGISTER_TYPE_UB
), 28),
308 brw_imm_uv(0x76543210));
310 /* A set bit in the pixel mask means the channel is enabled, but
311 * that is the opposite of gl_HelperInvocation so we need to invert
314 * The negate source-modifier bit of logical instructions on Gen8+
315 * performs 1's complement negation, so we can use that instead of
318 fs_reg inverted
= negate(shifted
);
319 if (v
->devinfo
->gen
< 8) {
320 inverted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
321 abld
.NOT(inverted
, shifted
);
324 /* We then resolve the 0/1 result to 0/~0 boolean values by ANDing
325 * with 1 and negating.
327 fs_reg anded
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
328 abld
.AND(anded
, inverted
, brw_imm_uw(1));
330 fs_reg dst
= abld
.vgrf(BRW_REGISTER_TYPE_D
, 1);
331 abld
.MOV(dst
, negate(retype(anded
, BRW_REGISTER_TYPE_D
)));
345 fs_visitor::nir_emit_system_values()
347 nir_system_values
= ralloc_array(mem_ctx
, fs_reg
, SYSTEM_VALUE_MAX
);
348 for (unsigned i
= 0; i
< SYSTEM_VALUE_MAX
; i
++) {
349 nir_system_values
[i
] = fs_reg();
352 nir_foreach_function(function
, nir
) {
353 assert(strcmp(function
->name
, "main") == 0);
354 assert(function
->impl
);
355 nir_foreach_block(block
, function
->impl
) {
356 emit_system_values_block(block
, this);
362 fs_visitor::nir_emit_impl(nir_function_impl
*impl
)
364 nir_locals
= ralloc_array(mem_ctx
, fs_reg
, impl
->reg_alloc
);
365 for (unsigned i
= 0; i
< impl
->reg_alloc
; i
++) {
366 nir_locals
[i
] = fs_reg();
369 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
370 unsigned array_elems
=
371 reg
->num_array_elems
== 0 ? 1 : reg
->num_array_elems
;
372 unsigned size
= array_elems
* reg
->num_components
;
373 const brw_reg_type reg_type
=
374 reg
->bit_size
== 32 ? BRW_REGISTER_TYPE_F
: BRW_REGISTER_TYPE_DF
;
375 nir_locals
[reg
->index
] = bld
.vgrf(reg_type
, size
);
378 nir_ssa_values
= reralloc(mem_ctx
, nir_ssa_values
, fs_reg
,
381 nir_emit_cf_list(&impl
->body
);
385 fs_visitor::nir_emit_cf_list(exec_list
*list
)
387 exec_list_validate(list
);
388 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
389 switch (node
->type
) {
391 nir_emit_if(nir_cf_node_as_if(node
));
394 case nir_cf_node_loop
:
395 nir_emit_loop(nir_cf_node_as_loop(node
));
398 case nir_cf_node_block
:
399 nir_emit_block(nir_cf_node_as_block(node
));
403 unreachable("Invalid CFG node block");
409 fs_visitor::nir_emit_if(nir_if
*if_stmt
)
411 /* first, put the condition into f0 */
412 fs_inst
*inst
= bld
.MOV(bld
.null_reg_d(),
413 retype(get_nir_src(if_stmt
->condition
),
414 BRW_REGISTER_TYPE_D
));
415 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
417 bld
.IF(BRW_PREDICATE_NORMAL
);
419 nir_emit_cf_list(&if_stmt
->then_list
);
421 /* note: if the else is empty, dead CF elimination will remove it */
422 bld
.emit(BRW_OPCODE_ELSE
);
424 nir_emit_cf_list(&if_stmt
->else_list
);
426 bld
.emit(BRW_OPCODE_ENDIF
);
430 fs_visitor::nir_emit_loop(nir_loop
*loop
)
432 bld
.emit(BRW_OPCODE_DO
);
434 nir_emit_cf_list(&loop
->body
);
436 bld
.emit(BRW_OPCODE_WHILE
);
440 fs_visitor::nir_emit_block(nir_block
*block
)
442 nir_foreach_instr(instr
, block
) {
443 nir_emit_instr(instr
);
448 fs_visitor::nir_emit_instr(nir_instr
*instr
)
450 const fs_builder abld
= bld
.annotate(NULL
, instr
);
452 switch (instr
->type
) {
453 case nir_instr_type_alu
:
454 nir_emit_alu(abld
, nir_instr_as_alu(instr
));
457 case nir_instr_type_intrinsic
:
459 case MESA_SHADER_VERTEX
:
460 nir_emit_vs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
462 case MESA_SHADER_TESS_CTRL
:
463 nir_emit_tcs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
465 case MESA_SHADER_TESS_EVAL
:
466 nir_emit_tes_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
468 case MESA_SHADER_GEOMETRY
:
469 nir_emit_gs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
471 case MESA_SHADER_FRAGMENT
:
472 nir_emit_fs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
474 case MESA_SHADER_COMPUTE
:
475 nir_emit_cs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
478 unreachable("unsupported shader stage");
482 case nir_instr_type_tex
:
483 nir_emit_texture(abld
, nir_instr_as_tex(instr
));
486 case nir_instr_type_load_const
:
487 nir_emit_load_const(abld
, nir_instr_as_load_const(instr
));
490 case nir_instr_type_ssa_undef
:
491 nir_emit_undef(abld
, nir_instr_as_ssa_undef(instr
));
494 case nir_instr_type_jump
:
495 nir_emit_jump(abld
, nir_instr_as_jump(instr
));
499 unreachable("unknown instruction type");
504 * Recognizes a parent instruction of nir_op_extract_* and changes the type to
508 fs_visitor::optimize_extract_to_float(nir_alu_instr
*instr
,
509 const fs_reg
&result
)
511 if (!instr
->src
[0].src
.is_ssa
||
512 !instr
->src
[0].src
.ssa
->parent_instr
)
515 if (instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
518 nir_alu_instr
*src0
=
519 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
521 if (src0
->op
!= nir_op_extract_u8
&& src0
->op
!= nir_op_extract_u16
&&
522 src0
->op
!= nir_op_extract_i8
&& src0
->op
!= nir_op_extract_i16
)
525 nir_const_value
*element
= nir_src_as_const_value(src0
->src
[1].src
);
526 assert(element
!= NULL
);
528 enum opcode extract_op
;
529 if (src0
->op
== nir_op_extract_u16
|| src0
->op
== nir_op_extract_i16
) {
530 assert(element
->u32
[0] <= 1);
531 extract_op
= SHADER_OPCODE_EXTRACT_WORD
;
533 assert(element
->u32
[0] <= 3);
534 extract_op
= SHADER_OPCODE_EXTRACT_BYTE
;
537 fs_reg op0
= get_nir_src(src0
->src
[0].src
);
538 op0
.type
= brw_type_for_nir_type(
539 (nir_alu_type
)(nir_op_infos
[src0
->op
].input_types
[0] |
540 nir_src_bit_size(src0
->src
[0].src
)));
541 op0
= offset(op0
, bld
, src0
->src
[0].swizzle
[0]);
543 set_saturate(instr
->dest
.saturate
,
544 bld
.emit(extract_op
, result
, op0
, brw_imm_ud(element
->u32
[0])));
549 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr
*instr
,
550 const fs_reg
&result
)
552 if (!instr
->src
[0].src
.is_ssa
||
553 instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_intrinsic
)
556 nir_intrinsic_instr
*src0
=
557 nir_instr_as_intrinsic(instr
->src
[0].src
.ssa
->parent_instr
);
559 if (src0
->intrinsic
!= nir_intrinsic_load_front_face
)
562 nir_const_value
*value1
= nir_src_as_const_value(instr
->src
[1].src
);
563 if (!value1
|| fabsf(value1
->f32
[0]) != 1.0f
)
566 nir_const_value
*value2
= nir_src_as_const_value(instr
->src
[2].src
);
567 if (!value2
|| fabsf(value2
->f32
[0]) != 1.0f
)
570 fs_reg tmp
= vgrf(glsl_type::int_type
);
572 if (devinfo
->gen
>= 6) {
573 /* Bit 15 of g0.0 is 0 if the polygon is front facing. */
574 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
576 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
578 * or(8) tmp.1<2>W g0.0<0,1,0>W 0x00003f80W
579 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
581 * and negate g0.0<0,1,0>W for (gl_FrontFacing ? -1.0 : 1.0).
583 * This negation looks like it's safe in practice, because bits 0:4 will
584 * surely be TRIANGLES
587 if (value1
->f32
[0] == -1.0f
) {
591 tmp
.type
= BRW_REGISTER_TYPE_W
;
592 tmp
.subreg_offset
= 2;
595 bld
.OR(tmp
, g0
, brw_imm_uw(0x3f80));
597 tmp
.type
= BRW_REGISTER_TYPE_D
;
598 tmp
.subreg_offset
= 0;
601 /* Bit 31 of g1.6 is 0 if the polygon is front facing. */
602 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
604 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
606 * or(8) tmp<1>D g1.6<0,1,0>D 0x3f800000D
607 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
609 * and negate g1.6<0,1,0>D for (gl_FrontFacing ? -1.0 : 1.0).
611 * This negation looks like it's safe in practice, because bits 0:4 will
612 * surely be TRIANGLES
615 if (value1
->f32
[0] == -1.0f
) {
619 bld
.OR(tmp
, g1_6
, brw_imm_d(0x3f800000));
621 bld
.AND(retype(result
, BRW_REGISTER_TYPE_D
), tmp
, brw_imm_d(0xbf800000));
627 fs_visitor::nir_emit_alu(const fs_builder
&bld
, nir_alu_instr
*instr
)
629 struct brw_wm_prog_key
*fs_key
= (struct brw_wm_prog_key
*) this->key
;
632 fs_reg result
= get_nir_dest(instr
->dest
.dest
);
633 result
.type
= brw_type_for_nir_type(
634 (nir_alu_type
)(nir_op_infos
[instr
->op
].output_type
|
635 nir_dest_bit_size(instr
->dest
.dest
)));
638 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
639 op
[i
] = get_nir_src(instr
->src
[i
].src
);
640 op
[i
].type
= brw_type_for_nir_type(
641 (nir_alu_type
)(nir_op_infos
[instr
->op
].input_types
[i
] |
642 nir_src_bit_size(instr
->src
[i
].src
)));
643 op
[i
].abs
= instr
->src
[i
].abs
;
644 op
[i
].negate
= instr
->src
[i
].negate
;
647 /* We get a bunch of mov's out of the from_ssa pass and they may still
648 * be vectorized. We'll handle them as a special-case. We'll also
649 * handle vecN here because it's basically the same thing.
657 fs_reg temp
= result
;
658 bool need_extra_copy
= false;
659 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
660 if (!instr
->src
[i
].src
.is_ssa
&&
661 instr
->dest
.dest
.reg
.reg
== instr
->src
[i
].src
.reg
.reg
) {
662 need_extra_copy
= true;
663 temp
= bld
.vgrf(result
.type
, 4);
668 for (unsigned i
= 0; i
< 4; i
++) {
669 if (!(instr
->dest
.write_mask
& (1 << i
)))
672 if (instr
->op
== nir_op_imov
|| instr
->op
== nir_op_fmov
) {
673 inst
= bld
.MOV(offset(temp
, bld
, i
),
674 offset(op
[0], bld
, instr
->src
[0].swizzle
[i
]));
676 inst
= bld
.MOV(offset(temp
, bld
, i
),
677 offset(op
[i
], bld
, instr
->src
[i
].swizzle
[0]));
679 inst
->saturate
= instr
->dest
.saturate
;
682 /* In this case the source and destination registers were the same,
683 * so we need to insert an extra set of moves in order to deal with
686 if (need_extra_copy
) {
687 for (unsigned i
= 0; i
< 4; i
++) {
688 if (!(instr
->dest
.write_mask
& (1 << i
)))
691 bld
.MOV(offset(result
, bld
, i
), offset(temp
, bld
, i
));
700 /* At this point, we have dealt with any instruction that operates on
701 * more than a single channel. Therefore, we can just adjust the source
702 * and destination registers for that channel and emit the instruction.
704 unsigned channel
= 0;
705 if (nir_op_infos
[instr
->op
].output_size
== 0) {
706 /* Since NIR is doing the scalarizing for us, we should only ever see
707 * vectorized operations with a single channel.
709 assert(_mesa_bitcount(instr
->dest
.write_mask
) == 1);
710 channel
= ffs(instr
->dest
.write_mask
) - 1;
712 result
= offset(result
, bld
, channel
);
715 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
716 assert(nir_op_infos
[instr
->op
].input_sizes
[i
] < 2);
717 op
[i
] = offset(op
[i
], bld
, instr
->src
[i
].swizzle
[channel
]);
723 if (optimize_extract_to_float(instr
, result
))
732 inst
= bld
.MOV(result
, op
[0]);
733 inst
->saturate
= instr
->dest
.saturate
;
738 bld
.MOV(result
, op
[0]);
742 if (type_sz(op
[0].type
) < 8) {
743 /* AND(val, 0x80000000) gives the sign bit.
745 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
748 bld
.CMP(bld
.null_reg_f(), op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
750 fs_reg result_int
= retype(result
, BRW_REGISTER_TYPE_UD
);
751 op
[0].type
= BRW_REGISTER_TYPE_UD
;
752 result
.type
= BRW_REGISTER_TYPE_UD
;
753 bld
.AND(result_int
, op
[0], brw_imm_ud(0x80000000u
));
755 inst
= bld
.OR(result_int
, result_int
, brw_imm_ud(0x3f800000u
));
756 inst
->predicate
= BRW_PREDICATE_NORMAL
;
757 if (instr
->dest
.saturate
) {
758 inst
= bld
.MOV(result
, result
);
759 inst
->saturate
= true;
762 /* For doubles we do the same but we need to consider:
764 * - 2-src instructions can't operate with 64-bit immediates
765 * - The sign is encoded in the high 32-bit of each DF
766 * - CMP with DF requires special handling in SIMD16
767 * - We need to produce a DF result.
770 /* 2-src instructions can't have 64-bit immediates, so put 0.0 in
771 * a register and compare with that.
773 fs_reg tmp
= vgrf(glsl_type::double_type
);
774 bld
.MOV(tmp
, brw_imm_df(0.0));
776 /* A direct DF CMP using the flag register (null dst) won't work in
777 * SIMD16 because the CMP will be split in two by lower_simd_width,
778 * resulting in two CMP instructions with the same dst (NULL),
779 * leading to dead code elimination of the first one. In SIMD8,
780 * however, there is no need to split the CMP and we can save some
783 fs_reg dst_tmp
= vgrf(glsl_type::double_type
);
784 bld
.CMP(dst_tmp
, op
[0], tmp
, BRW_CONDITIONAL_NZ
);
786 /* In SIMD16 we want to avoid using a NULL dst register with DF CMP,
787 * so we store the result of the comparison in a vgrf instead and
788 * then we generate a UD comparison from that that won't have to
789 * be split by lower_simd_width. This is what NIR does to handle
790 * double comparisons in the general case.
792 if (bld
.dispatch_width() == 16 ) {
793 fs_reg dst_tmp_ud
= retype(dst_tmp
, BRW_REGISTER_TYPE_UD
);
794 bld
.MOV(dst_tmp_ud
, subscript(dst_tmp
, BRW_REGISTER_TYPE_UD
, 0));
795 bld
.CMP(bld
.null_reg_ud(),
796 dst_tmp_ud
, brw_imm_ud(0), BRW_CONDITIONAL_NZ
);
799 /* Get the high 32-bit of each double component where the sign is */
800 fs_reg result_int
= retype(result
, BRW_REGISTER_TYPE_UD
);
801 bld
.MOV(result_int
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1));
803 /* Get the sign bit */
804 bld
.AND(result_int
, result_int
, brw_imm_ud(0x80000000u
));
806 /* Add 1.0 to the sign, predicated to skip the case of op[0] == 0.0 */
807 inst
= bld
.OR(result_int
, result_int
, brw_imm_ud(0x3f800000u
));
808 inst
->predicate
= BRW_PREDICATE_NORMAL
;
810 /* Convert from 32-bit float to 64-bit double */
811 result
.type
= BRW_REGISTER_TYPE_DF
;
812 inst
= bld
.MOV(result
, retype(result_int
, BRW_REGISTER_TYPE_F
));
814 if (instr
->dest
.saturate
) {
815 inst
= bld
.MOV(result
, result
);
816 inst
->saturate
= true;
823 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
824 * -> non-negative val generates 0x00000000.
825 * Predicated OR sets 1 if val is positive.
827 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
828 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_G
);
829 bld
.ASR(result
, op
[0], brw_imm_d(31));
830 inst
= bld
.OR(result
, result
, brw_imm_d(1));
831 inst
->predicate
= BRW_PREDICATE_NORMAL
;
835 inst
= bld
.emit(SHADER_OPCODE_RCP
, result
, op
[0]);
836 inst
->saturate
= instr
->dest
.saturate
;
840 inst
= bld
.emit(SHADER_OPCODE_EXP2
, result
, op
[0]);
841 inst
->saturate
= instr
->dest
.saturate
;
845 inst
= bld
.emit(SHADER_OPCODE_LOG2
, result
, op
[0]);
846 inst
->saturate
= instr
->dest
.saturate
;
850 inst
= bld
.emit(SHADER_OPCODE_SIN
, result
, op
[0]);
851 inst
->saturate
= instr
->dest
.saturate
;
855 inst
= bld
.emit(SHADER_OPCODE_COS
, result
, op
[0]);
856 inst
->saturate
= instr
->dest
.saturate
;
860 if (fs_key
->high_quality_derivatives
) {
861 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
863 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
865 inst
->saturate
= instr
->dest
.saturate
;
867 case nir_op_fddx_fine
:
868 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
869 inst
->saturate
= instr
->dest
.saturate
;
871 case nir_op_fddx_coarse
:
872 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
873 inst
->saturate
= instr
->dest
.saturate
;
876 if (fs_key
->high_quality_derivatives
) {
877 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0],
878 brw_imm_d(fs_key
->render_to_fbo
));
880 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0],
881 brw_imm_d(fs_key
->render_to_fbo
));
883 inst
->saturate
= instr
->dest
.saturate
;
885 case nir_op_fddy_fine
:
886 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0],
887 brw_imm_d(fs_key
->render_to_fbo
));
888 inst
->saturate
= instr
->dest
.saturate
;
890 case nir_op_fddy_coarse
:
891 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0],
892 brw_imm_d(fs_key
->render_to_fbo
));
893 inst
->saturate
= instr
->dest
.saturate
;
897 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
899 inst
= bld
.ADD(result
, op
[0], op
[1]);
900 inst
->saturate
= instr
->dest
.saturate
;
904 inst
= bld
.MUL(result
, op
[0], op
[1]);
905 inst
->saturate
= instr
->dest
.saturate
;
909 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
910 bld
.MUL(result
, op
[0], op
[1]);
913 case nir_op_imul_high
:
914 case nir_op_umul_high
:
915 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
916 bld
.emit(SHADER_OPCODE_MULH
, result
, op
[0], op
[1]);
921 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
922 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, result
, op
[0], op
[1]);
925 case nir_op_uadd_carry
:
926 unreachable("Should have been lowered by carry_to_arith().");
928 case nir_op_usub_borrow
:
929 unreachable("Should have been lowered by borrow_to_arith().");
933 /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
934 * appears that our hardware just does the right thing for signed
937 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
938 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
942 /* Get a regular C-style remainder. If a % b == 0, set the predicate. */
943 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
945 /* Math instructions don't support conditional mod */
946 inst
= bld
.MOV(bld
.null_reg_d(), result
);
947 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
949 /* Now, we need to determine if signs of the sources are different.
950 * When we XOR the sources, the top bit is 0 if they are the same and 1
951 * if they are different. We can then use a conditional modifier to
952 * turn that into a predicate. This leads us to an XOR.l instruction.
954 * Technically, according to the PRM, you're not allowed to use .l on a
955 * XOR instruction. However, emperical experiments and Curro's reading
956 * of the simulator source both indicate that it's safe.
958 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
959 inst
= bld
.XOR(tmp
, op
[0], op
[1]);
960 inst
->predicate
= BRW_PREDICATE_NORMAL
;
961 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
963 /* If the result of the initial remainder operation is non-zero and the
964 * two sources have different signs, add in a copy of op[1] to get the
965 * final integer modulus value.
967 inst
= bld
.ADD(result
, result
, op
[1]);
968 inst
->predicate
= BRW_PREDICATE_NORMAL
;
976 fs_reg dest
= result
;
977 if (nir_src_bit_size(instr
->src
[0].src
) > 32) {
978 dest
= bld
.vgrf(BRW_REGISTER_TYPE_DF
, 1);
980 brw_conditional_mod cond
;
983 cond
= BRW_CONDITIONAL_L
;
986 cond
= BRW_CONDITIONAL_GE
;
989 cond
= BRW_CONDITIONAL_Z
;
992 cond
= BRW_CONDITIONAL_NZ
;
995 unreachable("bad opcode");
997 bld
.CMP(dest
, op
[0], op
[1], cond
);
998 if (nir_src_bit_size(instr
->src
[0].src
) > 32) {
999 bld
.MOV(result
, subscript(dest
, BRW_REGISTER_TYPE_UD
, 0));
1006 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1007 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
1012 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1013 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
1017 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1018 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_Z
);
1022 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1023 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_NZ
);
1027 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1028 if (devinfo
->gen
>= 8) {
1029 op
[0] = resolve_source_modifiers(op
[0]);
1031 bld
.NOT(result
, op
[0]);
1034 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1035 if (devinfo
->gen
>= 8) {
1036 op
[0] = resolve_source_modifiers(op
[0]);
1037 op
[1] = resolve_source_modifiers(op
[1]);
1039 bld
.XOR(result
, op
[0], op
[1]);
1042 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1043 if (devinfo
->gen
>= 8) {
1044 op
[0] = resolve_source_modifiers(op
[0]);
1045 op
[1] = resolve_source_modifiers(op
[1]);
1047 bld
.OR(result
, op
[0], op
[1]);
1050 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1051 if (devinfo
->gen
>= 8) {
1052 op
[0] = resolve_source_modifiers(op
[0]);
1053 op
[1] = resolve_source_modifiers(op
[1]);
1055 bld
.AND(result
, op
[0], op
[1]);
1061 case nir_op_ball_fequal2
:
1062 case nir_op_ball_iequal2
:
1063 case nir_op_ball_fequal3
:
1064 case nir_op_ball_iequal3
:
1065 case nir_op_ball_fequal4
:
1066 case nir_op_ball_iequal4
:
1067 case nir_op_bany_fnequal2
:
1068 case nir_op_bany_inequal2
:
1069 case nir_op_bany_fnequal3
:
1070 case nir_op_bany_inequal3
:
1071 case nir_op_bany_fnequal4
:
1072 case nir_op_bany_inequal4
:
1073 unreachable("Lowered by nir_lower_alu_reductions");
1075 case nir_op_fnoise1_1
:
1076 case nir_op_fnoise1_2
:
1077 case nir_op_fnoise1_3
:
1078 case nir_op_fnoise1_4
:
1079 case nir_op_fnoise2_1
:
1080 case nir_op_fnoise2_2
:
1081 case nir_op_fnoise2_3
:
1082 case nir_op_fnoise2_4
:
1083 case nir_op_fnoise3_1
:
1084 case nir_op_fnoise3_2
:
1085 case nir_op_fnoise3_3
:
1086 case nir_op_fnoise3_4
:
1087 case nir_op_fnoise4_1
:
1088 case nir_op_fnoise4_2
:
1089 case nir_op_fnoise4_3
:
1090 case nir_op_fnoise4_4
:
1091 unreachable("not reached: should be handled by lower_noise");
1094 unreachable("not reached: should be handled by ldexp_to_arith()");
1097 inst
= bld
.emit(SHADER_OPCODE_SQRT
, result
, op
[0]);
1098 inst
->saturate
= instr
->dest
.saturate
;
1102 inst
= bld
.emit(SHADER_OPCODE_RSQ
, result
, op
[0]);
1103 inst
->saturate
= instr
->dest
.saturate
;
1108 bld
.MOV(result
, negate(op
[0]));
1112 bld
.CMP(result
, op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
1115 /* two-argument instructions can't take 64-bit immediates */
1116 fs_reg zero
= vgrf(glsl_type::double_type
);
1117 bld
.MOV(zero
, brw_imm_df(0.0));
1118 /* A SIMD16 execution needs to be split in two instructions, so use
1119 * a vgrf instead of the flag register as dst so instruction splitting
1122 fs_reg tmp
= vgrf(glsl_type::double_type
);
1123 bld
.CMP(tmp
, op
[0], zero
, BRW_CONDITIONAL_NZ
);
1124 bld
.MOV(result
, subscript(tmp
, BRW_REGISTER_TYPE_UD
, 0));
1128 bld
.CMP(result
, op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1132 inst
= bld
.RNDZ(result
, op
[0]);
1133 inst
->saturate
= instr
->dest
.saturate
;
1136 case nir_op_fceil
: {
1137 op
[0].negate
= !op
[0].negate
;
1138 fs_reg temp
= vgrf(glsl_type::float_type
);
1139 bld
.RNDD(temp
, op
[0]);
1141 inst
= bld
.MOV(result
, temp
);
1142 inst
->saturate
= instr
->dest
.saturate
;
1146 inst
= bld
.RNDD(result
, op
[0]);
1147 inst
->saturate
= instr
->dest
.saturate
;
1150 inst
= bld
.FRC(result
, op
[0]);
1151 inst
->saturate
= instr
->dest
.saturate
;
1153 case nir_op_fround_even
:
1154 inst
= bld
.RNDE(result
, op
[0]);
1155 inst
->saturate
= instr
->dest
.saturate
;
1158 case nir_op_fquantize2f16
: {
1159 fs_reg tmp16
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1160 fs_reg tmp32
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1161 fs_reg zero
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1163 /* The destination stride must be at least as big as the source stride. */
1164 tmp16
.type
= BRW_REGISTER_TYPE_W
;
1167 /* Check for denormal */
1168 fs_reg abs_src0
= op
[0];
1169 abs_src0
.abs
= true;
1170 bld
.CMP(bld
.null_reg_f(), abs_src0
, brw_imm_f(ldexpf(1.0, -14)),
1172 /* Get the appropriately signed zero */
1173 bld
.AND(retype(zero
, BRW_REGISTER_TYPE_UD
),
1174 retype(op
[0], BRW_REGISTER_TYPE_UD
),
1175 brw_imm_ud(0x80000000));
1176 /* Do the actual F32 -> F16 -> F32 conversion */
1177 bld
.emit(BRW_OPCODE_F32TO16
, tmp16
, op
[0]);
1178 bld
.emit(BRW_OPCODE_F16TO32
, tmp32
, tmp16
);
1179 /* Select that or zero based on normal status */
1180 inst
= bld
.SEL(result
, zero
, tmp32
);
1181 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1182 inst
->saturate
= instr
->dest
.saturate
;
1188 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1190 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
1191 inst
->saturate
= instr
->dest
.saturate
;
1196 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1198 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
1199 inst
->saturate
= instr
->dest
.saturate
;
1202 case nir_op_pack_snorm_2x16
:
1203 case nir_op_pack_snorm_4x8
:
1204 case nir_op_pack_unorm_2x16
:
1205 case nir_op_pack_unorm_4x8
:
1206 case nir_op_unpack_snorm_2x16
:
1207 case nir_op_unpack_snorm_4x8
:
1208 case nir_op_unpack_unorm_2x16
:
1209 case nir_op_unpack_unorm_4x8
:
1210 case nir_op_unpack_half_2x16
:
1211 case nir_op_pack_half_2x16
:
1212 unreachable("not reached: should be handled by lower_packing_builtins");
1214 case nir_op_unpack_half_2x16_split_x
:
1215 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
, result
, op
[0]);
1216 inst
->saturate
= instr
->dest
.saturate
;
1218 case nir_op_unpack_half_2x16_split_y
:
1219 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
, result
, op
[0]);
1220 inst
->saturate
= instr
->dest
.saturate
;
1223 case nir_op_pack_double_2x32_split
:
1224 /* Optimize the common case where we are re-packing a double with
1225 * the result of a previous double unpack. In this case we can take the
1226 * 32-bit value to use in the re-pack from the original double and bypass
1227 * the unpack operation.
1229 for (int i
= 0; i
< 2; i
++) {
1230 if (instr
->src
[i
].src
.is_ssa
)
1233 const nir_instr
*parent_instr
= instr
->src
[i
].src
.ssa
->parent_instr
;
1234 if (parent_instr
->type
== nir_instr_type_alu
)
1237 const nir_alu_instr
*alu_parent
= nir_instr_as_alu(parent_instr
);
1238 if (alu_parent
->op
== nir_op_unpack_double_2x32_split_x
||
1239 alu_parent
->op
== nir_op_unpack_double_2x32_split_y
)
1242 if (!alu_parent
->src
[0].src
.is_ssa
)
1245 op
[i
] = get_nir_src(alu_parent
->src
[0].src
);
1246 op
[i
] = offset(retype(op
[i
], BRW_REGISTER_TYPE_DF
), bld
,
1247 alu_parent
->src
[0].swizzle
[channel
]);
1248 if (alu_parent
->op
== nir_op_unpack_double_2x32_split_y
)
1249 op
[i
] = subscript(op
[i
], BRW_REGISTER_TYPE_UD
, 1);
1251 op
[i
] = subscript(op
[i
], BRW_REGISTER_TYPE_UD
, 0);
1253 bld
.emit(FS_OPCODE_PACK
, result
, op
[0], op
[1]);
1256 case nir_op_unpack_double_2x32_split_x
:
1257 case nir_op_unpack_double_2x32_split_y
: {
1258 /* Optimize the common case where we are unpacking from a double we have
1259 * previously packed. In this case we can just bypass the pack operation
1260 * and source directly from its arguments.
1262 unsigned index
= (instr
->op
== nir_op_unpack_double_2x32_split_x
) ? 0 : 1;
1263 if (instr
->src
[0].src
.is_ssa
) {
1264 nir_instr
*parent_instr
= instr
->src
[0].src
.ssa
->parent_instr
;
1265 if (parent_instr
->type
== nir_instr_type_alu
) {
1266 nir_alu_instr
*alu_parent
= nir_instr_as_alu(parent_instr
);
1267 if (alu_parent
->op
== nir_op_pack_double_2x32_split
&&
1268 alu_parent
->src
[index
].src
.is_ssa
) {
1269 op
[0] = retype(get_nir_src(alu_parent
->src
[index
].src
),
1270 BRW_REGISTER_TYPE_UD
);
1272 offset(op
[0], bld
, alu_parent
->src
[index
].swizzle
[channel
]);
1273 bld
.MOV(result
, op
[0]);
1279 if (instr
->op
== nir_op_unpack_double_2x32_split_x
)
1280 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 0));
1282 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1));
1287 inst
= bld
.emit(SHADER_OPCODE_POW
, result
, op
[0], op
[1]);
1288 inst
->saturate
= instr
->dest
.saturate
;
1291 case nir_op_bitfield_reverse
:
1292 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1293 bld
.BFREV(result
, op
[0]);
1296 case nir_op_bit_count
:
1297 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1298 bld
.CBIT(result
, op
[0]);
1301 case nir_op_ufind_msb
:
1302 case nir_op_ifind_msb
: {
1303 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1304 bld
.FBH(retype(result
, BRW_REGISTER_TYPE_UD
), op
[0]);
1306 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1307 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1308 * subtract the result from 31 to convert the MSB count into an LSB count.
1310 bld
.CMP(bld
.null_reg_d(), result
, brw_imm_d(-1), BRW_CONDITIONAL_NZ
);
1312 inst
= bld
.ADD(result
, result
, brw_imm_d(31));
1313 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1314 inst
->src
[0].negate
= true;
1318 case nir_op_find_lsb
:
1319 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1320 bld
.FBL(result
, op
[0]);
1323 case nir_op_ubitfield_extract
:
1324 case nir_op_ibitfield_extract
:
1325 unreachable("should have been lowered");
1328 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1329 bld
.BFE(result
, op
[2], op
[1], op
[0]);
1332 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1333 bld
.BFI1(result
, op
[0], op
[1]);
1336 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1337 bld
.BFI2(result
, op
[0], op
[1], op
[2]);
1340 case nir_op_bitfield_insert
:
1341 unreachable("not reached: should have been lowered");
1344 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1345 bld
.SHL(result
, op
[0], op
[1]);
1348 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1349 bld
.ASR(result
, op
[0], op
[1]);
1352 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1353 bld
.SHR(result
, op
[0], op
[1]);
1356 case nir_op_pack_half_2x16_split
:
1357 bld
.emit(FS_OPCODE_PACK_HALF_2x16_SPLIT
, result
, op
[0], op
[1]);
1361 inst
= bld
.MAD(result
, op
[2], op
[1], op
[0]);
1362 inst
->saturate
= instr
->dest
.saturate
;
1366 inst
= bld
.LRP(result
, op
[0], op
[1], op
[2]);
1367 inst
->saturate
= instr
->dest
.saturate
;
1371 if (optimize_frontfacing_ternary(instr
, result
))
1374 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1375 inst
= bld
.SEL(result
, op
[1], op
[2]);
1376 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1379 case nir_op_extract_u8
:
1380 case nir_op_extract_i8
: {
1381 nir_const_value
*byte
= nir_src_as_const_value(instr
->src
[1].src
);
1382 bld
.emit(SHADER_OPCODE_EXTRACT_BYTE
,
1383 result
, op
[0], brw_imm_ud(byte
->u32
[0]));
1387 case nir_op_extract_u16
:
1388 case nir_op_extract_i16
: {
1389 nir_const_value
*word
= nir_src_as_const_value(instr
->src
[1].src
);
1390 bld
.emit(SHADER_OPCODE_EXTRACT_WORD
,
1391 result
, op
[0], brw_imm_ud(word
->u32
[0]));
1396 unreachable("unhandled instruction");
1399 /* If we need to do a boolean resolve, replace the result with -(x & 1)
1400 * to sign extend the low bit to 0/~0
1402 if (devinfo
->gen
<= 5 &&
1403 (instr
->instr
.pass_flags
& BRW_NIR_BOOLEAN_MASK
) == BRW_NIR_BOOLEAN_NEEDS_RESOLVE
) {
1404 fs_reg masked
= vgrf(glsl_type::int_type
);
1405 bld
.AND(masked
, result
, brw_imm_d(1));
1406 masked
.negate
= true;
1407 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), masked
);
1412 fs_visitor::nir_emit_load_const(const fs_builder
&bld
,
1413 nir_load_const_instr
*instr
)
1415 const brw_reg_type reg_type
=
1416 instr
->def
.bit_size
== 32 ? BRW_REGISTER_TYPE_D
: BRW_REGISTER_TYPE_DF
;
1417 fs_reg reg
= bld
.vgrf(reg_type
, instr
->def
.num_components
);
1419 switch (instr
->def
.bit_size
) {
1421 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1422 bld
.MOV(offset(reg
, bld
, i
), brw_imm_d(instr
->value
.i32
[i
]));
1426 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1427 bld
.MOV(offset(reg
, bld
, i
), brw_imm_df(instr
->value
.f64
[i
]));
1431 unreachable("Invalid bit size");
1434 nir_ssa_values
[instr
->def
.index
] = reg
;
1438 fs_visitor::nir_emit_undef(const fs_builder
&bld
, nir_ssa_undef_instr
*instr
)
1440 const brw_reg_type reg_type
=
1441 instr
->def
.bit_size
== 32 ? BRW_REGISTER_TYPE_D
: BRW_REGISTER_TYPE_DF
;
1442 nir_ssa_values
[instr
->def
.index
] =
1443 bld
.vgrf(reg_type
, instr
->def
.num_components
);
1447 fs_visitor::get_nir_src(nir_src src
)
1451 reg
= nir_ssa_values
[src
.ssa
->index
];
1453 /* We don't handle indirects on locals */
1454 assert(src
.reg
.indirect
== NULL
);
1455 reg
= offset(nir_locals
[src
.reg
.reg
->index
], bld
,
1456 src
.reg
.base_offset
* src
.reg
.reg
->num_components
);
1459 /* to avoid floating-point denorm flushing problems, set the type by
1460 * default to D - instructions that need floating point semantics will set
1461 * this to F if they need to
1463 return retype(reg
, BRW_REGISTER_TYPE_D
);
1467 fs_visitor::get_nir_dest(nir_dest dest
)
1470 const brw_reg_type reg_type
=
1471 dest
.ssa
.bit_size
== 32 ? BRW_REGISTER_TYPE_F
: BRW_REGISTER_TYPE_DF
;
1472 nir_ssa_values
[dest
.ssa
.index
] =
1473 bld
.vgrf(reg_type
, dest
.ssa
.num_components
);
1474 return nir_ssa_values
[dest
.ssa
.index
];
1476 /* We don't handle indirects on locals */
1477 assert(dest
.reg
.indirect
== NULL
);
1478 return offset(nir_locals
[dest
.reg
.reg
->index
], bld
,
1479 dest
.reg
.base_offset
* dest
.reg
.reg
->num_components
);
1484 fs_visitor::get_nir_image_deref(const nir_deref_var
*deref
)
1486 fs_reg
image(UNIFORM
, deref
->var
->data
.driver_location
/ 4,
1487 BRW_REGISTER_TYPE_UD
);
1489 unsigned indirect_max
= 0;
1491 for (const nir_deref
*tail
= &deref
->deref
; tail
->child
;
1492 tail
= tail
->child
) {
1493 const nir_deref_array
*deref_array
= nir_deref_as_array(tail
->child
);
1494 assert(tail
->child
->deref_type
== nir_deref_type_array
);
1495 const unsigned size
= glsl_get_length(tail
->type
);
1496 const unsigned element_size
= type_size_scalar(deref_array
->deref
.type
);
1497 const unsigned base
= MIN2(deref_array
->base_offset
, size
- 1);
1498 image
= offset(image
, bld
, base
* element_size
);
1500 if (deref_array
->deref_array_type
== nir_deref_array_type_indirect
) {
1501 fs_reg tmp
= vgrf(glsl_type::uint_type
);
1503 /* Accessing an invalid surface index with the dataport can result
1504 * in a hang. According to the spec "if the index used to
1505 * select an individual element is negative or greater than or
1506 * equal to the size of the array, the results of the operation
1507 * are undefined but may not lead to termination" -- which is one
1508 * of the possible outcomes of the hang. Clamp the index to
1509 * prevent access outside of the array bounds.
1511 bld
.emit_minmax(tmp
, retype(get_nir_src(deref_array
->indirect
),
1512 BRW_REGISTER_TYPE_UD
),
1513 brw_imm_ud(size
- base
- 1), BRW_CONDITIONAL_L
);
1515 indirect_max
+= element_size
* (tail
->type
->length
- 1);
1517 bld
.MUL(tmp
, tmp
, brw_imm_ud(element_size
* 4));
1518 if (indirect
.file
== BAD_FILE
) {
1521 bld
.ADD(indirect
, indirect
, tmp
);
1526 if (indirect
.file
== BAD_FILE
) {
1529 /* Emit a pile of MOVs to load the uniform into a temporary. The
1530 * dead-code elimination pass will get rid of what we don't use.
1532 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, BRW_IMAGE_PARAM_SIZE
);
1533 for (unsigned j
= 0; j
< BRW_IMAGE_PARAM_SIZE
; j
++) {
1534 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
1535 offset(tmp
, bld
, j
), offset(image
, bld
, j
),
1536 indirect
, brw_imm_ud((indirect_max
+ 1) * 4));
1543 fs_visitor::emit_percomp(const fs_builder
&bld
, const fs_inst
&inst
,
1546 for (unsigned i
= 0; i
< 4; i
++) {
1547 if (!((wr_mask
>> i
) & 1))
1550 fs_inst
*new_inst
= new(mem_ctx
) fs_inst(inst
);
1551 new_inst
->dst
= offset(new_inst
->dst
, bld
, i
);
1552 for (unsigned j
= 0; j
< new_inst
->sources
; j
++)
1553 if (new_inst
->src
[j
].file
== VGRF
)
1554 new_inst
->src
[j
] = offset(new_inst
->src
[j
], bld
, i
);
1561 * Get the matching channel register datatype for an image intrinsic of the
1562 * specified GLSL image type.
1565 get_image_base_type(const glsl_type
*type
)
1567 switch ((glsl_base_type
)type
->sampled_type
) {
1568 case GLSL_TYPE_UINT
:
1569 return BRW_REGISTER_TYPE_UD
;
1571 return BRW_REGISTER_TYPE_D
;
1572 case GLSL_TYPE_FLOAT
:
1573 return BRW_REGISTER_TYPE_F
;
1575 unreachable("Not reached.");
1580 * Get the appropriate atomic op for an image atomic intrinsic.
1583 get_image_atomic_op(nir_intrinsic_op op
, const glsl_type
*type
)
1586 case nir_intrinsic_image_atomic_add
:
1588 case nir_intrinsic_image_atomic_min
:
1589 return (get_image_base_type(type
) == BRW_REGISTER_TYPE_D
?
1590 BRW_AOP_IMIN
: BRW_AOP_UMIN
);
1591 case nir_intrinsic_image_atomic_max
:
1592 return (get_image_base_type(type
) == BRW_REGISTER_TYPE_D
?
1593 BRW_AOP_IMAX
: BRW_AOP_UMAX
);
1594 case nir_intrinsic_image_atomic_and
:
1596 case nir_intrinsic_image_atomic_or
:
1598 case nir_intrinsic_image_atomic_xor
:
1600 case nir_intrinsic_image_atomic_exchange
:
1602 case nir_intrinsic_image_atomic_comp_swap
:
1603 return BRW_AOP_CMPWR
;
1605 unreachable("Not reachable.");
1610 emit_pixel_interpolater_send(const fs_builder
&bld
,
1615 glsl_interp_qualifier interpolation
)
1621 if (src
.file
== BAD_FILE
) {
1623 payload
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 1);
1627 mlen
= 2 * bld
.dispatch_width() / 8;
1630 inst
= bld
.emit(opcode
, dst
, payload
, desc
);
1632 /* 2 floats per slot returned */
1633 inst
->regs_written
= 2 * bld
.dispatch_width() / 8;
1634 inst
->pi_noperspective
= interpolation
== INTERP_QUALIFIER_NOPERSPECTIVE
;
1640 * Computes 1 << x, given a D/UD register containing some value x.
1643 intexp2(const fs_builder
&bld
, const fs_reg
&x
)
1645 assert(x
.type
== BRW_REGISTER_TYPE_UD
|| x
.type
== BRW_REGISTER_TYPE_D
);
1647 fs_reg result
= bld
.vgrf(x
.type
, 1);
1648 fs_reg one
= bld
.vgrf(x
.type
, 1);
1650 bld
.MOV(one
, retype(brw_imm_d(1), one
.type
));
1651 bld
.SHL(result
, one
, x
);
1656 fs_visitor::emit_gs_end_primitive(const nir_src
&vertex_count_nir_src
)
1658 assert(stage
== MESA_SHADER_GEOMETRY
);
1660 struct brw_gs_prog_data
*gs_prog_data
=
1661 (struct brw_gs_prog_data
*) prog_data
;
1663 /* We can only do EndPrimitive() functionality when the control data
1664 * consists of cut bits. Fortunately, the only time it isn't is when the
1665 * output type is points, in which case EndPrimitive() is a no-op.
1667 if (gs_prog_data
->control_data_format
!=
1668 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
) {
1672 /* Cut bits use one bit per vertex. */
1673 assert(gs_compile
->control_data_bits_per_vertex
== 1);
1675 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1676 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
1678 /* Cut bit n should be set to 1 if EndPrimitive() was called after emitting
1679 * vertex n, 0 otherwise. So all we need to do here is mark bit
1680 * (vertex_count - 1) % 32 in the cut_bits register to indicate that
1681 * EndPrimitive() was called after emitting vertex (vertex_count - 1);
1682 * vec4_gs_visitor::emit_control_data_bits() will take care of the rest.
1684 * Note that if EndPrimitive() is called before emitting any vertices, this
1685 * will cause us to set bit 31 of the control_data_bits register to 1.
1686 * That's fine because:
1688 * - If max_vertices < 32, then vertex number 31 (zero-based) will never be
1689 * output, so the hardware will ignore cut bit 31.
1691 * - If max_vertices == 32, then vertex number 31 is guaranteed to be the
1692 * last vertex, so setting cut bit 31 has no effect (since the primitive
1693 * is automatically ended when the GS terminates).
1695 * - If max_vertices > 32, then the ir_emit_vertex visitor will reset the
1696 * control_data_bits register to 0 when the first vertex is emitted.
1699 const fs_builder abld
= bld
.annotate("end primitive");
1701 /* control_data_bits |= 1 << ((vertex_count - 1) % 32) */
1702 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1703 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1704 fs_reg mask
= intexp2(abld
, prev_count
);
1705 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1706 * attention to the lower 5 bits of its second source argument, so on this
1707 * architecture, 1 << (vertex_count - 1) is equivalent to 1 <<
1708 * ((vertex_count - 1) % 32).
1710 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1714 fs_visitor::emit_gs_control_data_bits(const fs_reg
&vertex_count
)
1716 assert(stage
== MESA_SHADER_GEOMETRY
);
1717 assert(gs_compile
->control_data_bits_per_vertex
!= 0);
1719 struct brw_gs_prog_data
*gs_prog_data
=
1720 (struct brw_gs_prog_data
*) prog_data
;
1722 const fs_builder abld
= bld
.annotate("emit control data bits");
1723 const fs_builder fwa_bld
= bld
.exec_all();
1725 /* We use a single UD register to accumulate control data bits (32 bits
1726 * for each of the SIMD8 channels). So we need to write a DWord (32 bits)
1729 * Unfortunately, the URB_WRITE_SIMD8 message uses 128-bit (OWord) offsets.
1730 * We have select a 128-bit group via the Global and Per-Slot Offsets, then
1731 * use the Channel Mask phase to enable/disable which DWord within that
1732 * group to write. (Remember, different SIMD8 channels may have emitted
1733 * different numbers of vertices, so we may need per-slot offsets.)
1735 * Channel masking presents an annoying problem: we may have to replicate
1736 * the data up to 4 times:
1738 * Msg = Handles, Per-Slot Offsets, Channel Masks, Data, Data, Data, Data.
1740 * To avoid penalizing shaders that emit a small number of vertices, we
1741 * can avoid these sometimes: if the size of the control data header is
1742 * <= 128 bits, then there is only 1 OWord. All SIMD8 channels will land
1743 * land in the same 128-bit group, so we can skip per-slot offsets.
1745 * Similarly, if the control data header is <= 32 bits, there is only one
1746 * DWord, so we can skip channel masks.
1748 enum opcode opcode
= SHADER_OPCODE_URB_WRITE_SIMD8
;
1750 fs_reg channel_mask
, per_slot_offset
;
1752 if (gs_compile
->control_data_header_size_bits
> 32) {
1753 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
1754 channel_mask
= vgrf(glsl_type::uint_type
);
1757 if (gs_compile
->control_data_header_size_bits
> 128) {
1758 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
;
1759 per_slot_offset
= vgrf(glsl_type::uint_type
);
1762 /* Figure out which DWord we're trying to write to using the formula:
1764 * dword_index = (vertex_count - 1) * bits_per_vertex / 32
1766 * Since bits_per_vertex is a power of two, and is known at compile
1767 * time, this can be optimized to:
1769 * dword_index = (vertex_count - 1) >> (6 - log2(bits_per_vertex))
1771 if (opcode
!= SHADER_OPCODE_URB_WRITE_SIMD8
) {
1772 fs_reg dword_index
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1773 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1774 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1775 unsigned log2_bits_per_vertex
=
1776 _mesa_fls(gs_compile
->control_data_bits_per_vertex
);
1777 abld
.SHR(dword_index
, prev_count
, brw_imm_ud(6u - log2_bits_per_vertex
));
1779 if (per_slot_offset
.file
!= BAD_FILE
) {
1780 /* Set the per-slot offset to dword_index / 4, so that we'll write to
1781 * the appropriate OWord within the control data header.
1783 abld
.SHR(per_slot_offset
, dword_index
, brw_imm_ud(2u));
1786 /* Set the channel masks to 1 << (dword_index % 4), so that we'll
1787 * write to the appropriate DWORD within the OWORD.
1789 fs_reg channel
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1790 fwa_bld
.AND(channel
, dword_index
, brw_imm_ud(3u));
1791 channel_mask
= intexp2(fwa_bld
, channel
);
1792 /* Then the channel masks need to be in bits 23:16. */
1793 fwa_bld
.SHL(channel_mask
, channel_mask
, brw_imm_ud(16u));
1796 /* Store the control data bits in the message payload and send it. */
1798 if (channel_mask
.file
!= BAD_FILE
)
1799 mlen
+= 4; /* channel masks, plus 3 extra copies of the data */
1800 if (per_slot_offset
.file
!= BAD_FILE
)
1803 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
1804 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, mlen
);
1806 sources
[i
++] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1807 if (per_slot_offset
.file
!= BAD_FILE
)
1808 sources
[i
++] = per_slot_offset
;
1809 if (channel_mask
.file
!= BAD_FILE
)
1810 sources
[i
++] = channel_mask
;
1812 sources
[i
++] = this->control_data_bits
;
1815 abld
.LOAD_PAYLOAD(payload
, sources
, mlen
, mlen
);
1816 fs_inst
*inst
= abld
.emit(opcode
, reg_undef
, payload
);
1818 /* We need to increment Global Offset by 256-bits to make room for
1819 * Broadwell's extra "Vertex Count" payload at the beginning of the
1820 * URB entry. Since this is an OWord message, Global Offset is counted
1821 * in 128-bit units, so we must set it to 2.
1823 if (gs_prog_data
->static_vertex_count
== -1)
1828 fs_visitor::set_gs_stream_control_data_bits(const fs_reg
&vertex_count
,
1831 /* control_data_bits |= stream_id << ((2 * (vertex_count - 1)) % 32) */
1833 /* Note: we are calling this *before* increasing vertex_count, so
1834 * this->vertex_count == vertex_count - 1 in the formula above.
1837 /* Stream mode uses 2 bits per vertex */
1838 assert(gs_compile
->control_data_bits_per_vertex
== 2);
1840 /* Must be a valid stream */
1841 assert(stream_id
>= 0 && stream_id
< MAX_VERTEX_STREAMS
);
1843 /* Control data bits are initialized to 0 so we don't have to set any
1844 * bits when sending vertices to stream 0.
1849 const fs_builder abld
= bld
.annotate("set stream control data bits", NULL
);
1851 /* reg::sid = stream_id */
1852 fs_reg sid
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1853 abld
.MOV(sid
, brw_imm_ud(stream_id
));
1855 /* reg:shift_count = 2 * (vertex_count - 1) */
1856 fs_reg shift_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1857 abld
.SHL(shift_count
, vertex_count
, brw_imm_ud(1u));
1859 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1860 * attention to the lower 5 bits of its second source argument, so on this
1861 * architecture, stream_id << 2 * (vertex_count - 1) is equivalent to
1862 * stream_id << ((2 * (vertex_count - 1)) % 32).
1864 fs_reg mask
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1865 abld
.SHL(mask
, sid
, shift_count
);
1866 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1870 fs_visitor::emit_gs_vertex(const nir_src
&vertex_count_nir_src
,
1873 assert(stage
== MESA_SHADER_GEOMETRY
);
1875 struct brw_gs_prog_data
*gs_prog_data
=
1876 (struct brw_gs_prog_data
*) prog_data
;
1878 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1879 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
1881 /* Haswell and later hardware ignores the "Render Stream Select" bits
1882 * from the 3DSTATE_STREAMOUT packet when the SOL stage is disabled,
1883 * and instead sends all primitives down the pipeline for rasterization.
1884 * If the SOL stage is enabled, "Render Stream Select" is honored and
1885 * primitives bound to non-zero streams are discarded after stream output.
1887 * Since the only purpose of primives sent to non-zero streams is to
1888 * be recorded by transform feedback, we can simply discard all geometry
1889 * bound to these streams when transform feedback is disabled.
1891 if (stream_id
> 0 && !nir
->info
.has_transform_feedback_varyings
)
1894 /* If we're outputting 32 control data bits or less, then we can wait
1895 * until the shader is over to output them all. Otherwise we need to
1896 * output them as we go. Now is the time to do it, since we're about to
1897 * output the vertex_count'th vertex, so it's guaranteed that the
1898 * control data bits associated with the (vertex_count - 1)th vertex are
1901 if (gs_compile
->control_data_header_size_bits
> 32) {
1902 const fs_builder abld
=
1903 bld
.annotate("emit vertex: emit control data bits");
1905 /* Only emit control data bits if we've finished accumulating a batch
1906 * of 32 bits. This is the case when:
1908 * (vertex_count * bits_per_vertex) % 32 == 0
1910 * (in other words, when the last 5 bits of vertex_count *
1911 * bits_per_vertex are 0). Assuming bits_per_vertex == 2^n for some
1912 * integer n (which is always the case, since bits_per_vertex is
1913 * always 1 or 2), this is equivalent to requiring that the last 5-n
1914 * bits of vertex_count are 0:
1916 * vertex_count & (2^(5-n) - 1) == 0
1918 * 2^(5-n) == 2^5 / 2^n == 32 / bits_per_vertex, so this is
1921 * vertex_count & (32 / bits_per_vertex - 1) == 0
1923 * TODO: If vertex_count is an immediate, we could do some of this math
1924 * at compile time...
1927 abld
.AND(bld
.null_reg_d(), vertex_count
,
1928 brw_imm_ud(32u / gs_compile
->control_data_bits_per_vertex
- 1u));
1929 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1931 abld
.IF(BRW_PREDICATE_NORMAL
);
1932 /* If vertex_count is 0, then no control data bits have been
1933 * accumulated yet, so we can skip emitting them.
1935 abld
.CMP(bld
.null_reg_d(), vertex_count
, brw_imm_ud(0u),
1936 BRW_CONDITIONAL_NEQ
);
1937 abld
.IF(BRW_PREDICATE_NORMAL
);
1938 emit_gs_control_data_bits(vertex_count
);
1939 abld
.emit(BRW_OPCODE_ENDIF
);
1941 /* Reset control_data_bits to 0 so we can start accumulating a new
1944 * Note: in the case where vertex_count == 0, this neutralizes the
1945 * effect of any call to EndPrimitive() that the shader may have
1946 * made before outputting its first vertex.
1948 inst
= abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
1949 inst
->force_writemask_all
= true;
1950 abld
.emit(BRW_OPCODE_ENDIF
);
1953 emit_urb_writes(vertex_count
);
1955 /* In stream mode we have to set control data bits for all vertices
1956 * unless we have disabled control data bits completely (which we do
1957 * do for GL_POINTS outputs that don't use streams).
1959 if (gs_compile
->control_data_header_size_bits
> 0 &&
1960 gs_prog_data
->control_data_format
==
1961 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
) {
1962 set_gs_stream_control_data_bits(vertex_count
, stream_id
);
1967 fs_visitor::emit_gs_input_load(const fs_reg
&dst
,
1968 const nir_src
&vertex_src
,
1969 unsigned base_offset
,
1970 const nir_src
&offset_src
,
1971 unsigned num_components
)
1973 struct brw_gs_prog_data
*gs_prog_data
= (struct brw_gs_prog_data
*) prog_data
;
1975 nir_const_value
*vertex_const
= nir_src_as_const_value(vertex_src
);
1976 nir_const_value
*offset_const
= nir_src_as_const_value(offset_src
);
1977 const unsigned push_reg_count
= gs_prog_data
->base
.urb_read_length
* 8;
1979 /* Offset 0 is the VUE header, which contains VARYING_SLOT_LAYER [.y],
1980 * VARYING_SLOT_VIEWPORT [.z], and VARYING_SLOT_PSIZ [.w]. Only
1981 * gl_PointSize is available as a GS input, however, so it must be that.
1983 const bool is_point_size
= (base_offset
== 0);
1985 /* TODO: figure out push input layout for invocations == 1 */
1986 if (gs_prog_data
->invocations
== 1 &&
1987 offset_const
!= NULL
&& vertex_const
!= NULL
&&
1988 4 * (base_offset
+ offset_const
->u32
[0]) < push_reg_count
) {
1989 int imm_offset
= (base_offset
+ offset_const
->u32
[0]) * 4 +
1990 vertex_const
->u32
[0] * push_reg_count
;
1991 /* This input was pushed into registers. */
1992 if (is_point_size
) {
1993 /* gl_PointSize comes in .w */
1994 bld
.MOV(dst
, fs_reg(ATTR
, imm_offset
+ 3, dst
.type
));
1996 for (unsigned i
= 0; i
< num_components
; i
++) {
1997 bld
.MOV(offset(dst
, bld
, i
),
1998 fs_reg(ATTR
, imm_offset
+ i
, dst
.type
));
2004 /* Resort to the pull model. Ensure the VUE handles are provided. */
2005 gs_prog_data
->base
.include_vue_handles
= true;
2007 unsigned first_icp_handle
= gs_prog_data
->include_primitive_id
? 3 : 2;
2008 fs_reg icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2010 if (gs_prog_data
->invocations
== 1) {
2012 /* The vertex index is constant; just select the proper URB handle. */
2014 retype(brw_vec8_grf(first_icp_handle
+ vertex_const
->i32
[0], 0),
2015 BRW_REGISTER_TYPE_UD
);
2017 /* The vertex index is non-constant. We need to use indirect
2018 * addressing to fetch the proper URB handle.
2020 * First, we start with the sequence <7, 6, 5, 4, 3, 2, 1, 0>
2021 * indicating that channel <n> should read the handle from
2022 * DWord <n>. We convert that to bytes by multiplying by 4.
2024 * Next, we convert the vertex index to bytes by multiplying
2025 * by 32 (shifting by 5), and add the two together. This is
2026 * the final indirect byte offset.
2028 fs_reg sequence
= bld
.vgrf(BRW_REGISTER_TYPE_W
, 1);
2029 fs_reg channel_offsets
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2030 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2031 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2033 /* sequence = <7, 6, 5, 4, 3, 2, 1, 0> */
2034 bld
.MOV(sequence
, fs_reg(brw_imm_v(0x76543210)));
2035 /* channel_offsets = 4 * sequence = <28, 24, 20, 16, 12, 8, 4, 0> */
2036 bld
.SHL(channel_offsets
, sequence
, brw_imm_ud(2u));
2037 /* Convert vertex_index to bytes (multiply by 32) */
2038 bld
.SHL(vertex_offset_bytes
,
2039 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2041 bld
.ADD(icp_offset_bytes
, vertex_offset_bytes
, channel_offsets
);
2043 /* Use first_icp_handle as the base offset. There is one register
2044 * of URB handles per vertex, so inform the register allocator that
2045 * we might read up to nir->info.gs.vertices_in registers.
2047 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2048 fs_reg(brw_vec8_grf(first_icp_handle
, 0)),
2049 fs_reg(icp_offset_bytes
),
2050 brw_imm_ud(nir
->info
.gs
.vertices_in
* REG_SIZE
));
2053 assert(gs_prog_data
->invocations
> 1);
2056 assert(devinfo
->gen
>= 9 || vertex_const
->i32
[0] <= 5);
2058 retype(brw_vec1_grf(first_icp_handle
+
2059 vertex_const
->i32
[0] / 8,
2060 vertex_const
->i32
[0] % 8),
2061 BRW_REGISTER_TYPE_UD
));
2063 /* The vertex index is non-constant. We need to use indirect
2064 * addressing to fetch the proper URB handle.
2067 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2069 /* Convert vertex_index to bytes (multiply by 4) */
2070 bld
.SHL(icp_offset_bytes
,
2071 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2074 /* Use first_icp_handle as the base offset. There is one DWord
2075 * of URB handles per vertex, so inform the register allocator that
2076 * we might read up to ceil(nir->info.gs.vertices_in / 8) registers.
2078 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2079 fs_reg(brw_vec8_grf(first_icp_handle
, 0)),
2080 fs_reg(icp_offset_bytes
),
2081 brw_imm_ud(DIV_ROUND_UP(nir
->info
.gs
.vertices_in
, 8) *
2088 /* Constant indexing - use global offset. */
2089 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2090 inst
->offset
= base_offset
+ offset_const
->u32
[0];
2091 inst
->base_mrf
= -1;
2093 inst
->regs_written
= num_components
;
2095 /* Indirect indexing - use per-slot offsets as well. */
2096 const fs_reg srcs
[] = { icp_handle
, get_nir_src(offset_src
) };
2097 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2098 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2100 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2101 inst
->offset
= base_offset
;
2102 inst
->base_mrf
= -1;
2104 inst
->regs_written
= num_components
;
2107 if (is_point_size
) {
2108 /* Read the whole VUE header (because of alignment) and read .w. */
2109 fs_reg tmp
= bld
.vgrf(dst
.type
, 4);
2111 inst
->regs_written
= 4;
2112 bld
.MOV(dst
, offset(tmp
, bld
, 3));
2117 fs_visitor::get_indirect_offset(nir_intrinsic_instr
*instr
)
2119 nir_src
*offset_src
= nir_get_io_offset_src(instr
);
2120 nir_const_value
*const_value
= nir_src_as_const_value(*offset_src
);
2123 /* The only constant offset we should find is 0. brw_nir.c's
2124 * add_const_offset_to_base() will fold other constant offsets
2125 * into instr->const_index[0].
2127 assert(const_value
->u32
[0] == 0);
2131 return get_nir_src(*offset_src
);
2135 do_untyped_vector_read(const fs_builder
&bld
,
2137 const fs_reg surf_index
,
2138 const fs_reg offset_reg
,
2139 unsigned num_components
)
2141 if (type_sz(dest
.type
) == 4) {
2142 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, offset_reg
,
2145 BRW_PREDICATE_NONE
);
2146 read_result
.type
= dest
.type
;
2147 for (unsigned i
= 0; i
< num_components
; i
++)
2148 bld
.MOV(offset(dest
, bld
, i
), offset(read_result
, bld
, i
));
2149 } else if (type_sz(dest
.type
) == 8) {
2150 /* Reading a dvec, so we need to:
2152 * 1. Multiply num_components by 2, to account for the fact that we
2153 * need to read 64-bit components.
2154 * 2. Shuffle the result of the load to form valid 64-bit elements
2155 * 3. Emit a second load (for components z/w) if needed.
2157 fs_reg read_offset
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
2158 bld
.MOV(read_offset
, offset_reg
);
2160 int iters
= num_components
<= 2 ? 1 : 2;
2162 /* Load the dvec, the first iteration loads components x/y, the second
2163 * iteration, if needed, loads components z/w
2165 for (int it
= 0; it
< iters
; it
++) {
2166 /* Compute number of components to read in this iteration */
2167 int iter_components
= MIN2(2, num_components
);
2168 num_components
-= iter_components
;
2170 /* Read. Since this message reads 32-bit components, we need to
2171 * read twice as many components.
2173 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, read_offset
,
2175 iter_components
* 2,
2176 BRW_PREDICATE_NONE
);
2178 /* Shuffle the 32-bit load result into valid 64-bit data */
2179 const fs_reg packed_result
= bld
.vgrf(dest
.type
, iter_components
);
2180 shuffle_32bit_load_result_to_64bit_data(
2181 bld
, packed_result
, read_result
, iter_components
);
2183 /* Move each component to its destination */
2184 read_result
= retype(read_result
, BRW_REGISTER_TYPE_DF
);
2185 for (int c
= 0; c
< iter_components
; c
++) {
2186 bld
.MOV(offset(dest
, bld
, it
* 2 + c
),
2187 offset(packed_result
, bld
, c
));
2190 bld
.ADD(read_offset
, read_offset
, brw_imm_ud(16));
2193 unreachable("Unsupported type");
2198 fs_visitor::nir_emit_vs_intrinsic(const fs_builder
&bld
,
2199 nir_intrinsic_instr
*instr
)
2201 assert(stage
== MESA_SHADER_VERTEX
);
2204 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2205 dest
= get_nir_dest(instr
->dest
);
2207 switch (instr
->intrinsic
) {
2208 case nir_intrinsic_load_vertex_id
:
2209 unreachable("should be lowered by lower_vertex_id()");
2211 case nir_intrinsic_load_vertex_id_zero_base
:
2212 case nir_intrinsic_load_base_vertex
:
2213 case nir_intrinsic_load_instance_id
:
2214 case nir_intrinsic_load_base_instance
:
2215 case nir_intrinsic_load_draw_id
: {
2216 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
2217 fs_reg val
= nir_system_values
[sv
];
2218 assert(val
.file
!= BAD_FILE
);
2219 dest
.type
= val
.type
;
2225 nir_emit_intrinsic(bld
, instr
);
2231 fs_visitor::nir_emit_tcs_intrinsic(const fs_builder
&bld
,
2232 nir_intrinsic_instr
*instr
)
2234 assert(stage
== MESA_SHADER_TESS_CTRL
);
2235 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
2236 struct brw_tcs_prog_data
*tcs_prog_data
=
2237 (struct brw_tcs_prog_data
*) prog_data
;
2240 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2241 dst
= get_nir_dest(instr
->dest
);
2243 switch (instr
->intrinsic
) {
2244 case nir_intrinsic_load_primitive_id
:
2245 bld
.MOV(dst
, fs_reg(brw_vec1_grf(0, 1)));
2247 case nir_intrinsic_load_invocation_id
:
2248 bld
.MOV(retype(dst
, invocation_id
.type
), invocation_id
);
2250 case nir_intrinsic_load_patch_vertices_in
:
2251 bld
.MOV(retype(dst
, BRW_REGISTER_TYPE_D
),
2252 brw_imm_d(tcs_key
->input_vertices
));
2255 case nir_intrinsic_barrier
: {
2256 if (tcs_prog_data
->instances
== 1)
2259 fs_reg m0
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2260 fs_reg m0_2
= byte_offset(m0
, 2 * sizeof(uint32_t));
2262 const fs_builder fwa_bld
= bld
.exec_all();
2264 /* Zero the message header */
2265 fwa_bld
.MOV(m0
, brw_imm_ud(0u));
2267 /* Copy "Barrier ID" from r0.2, bits 16:13 */
2268 fwa_bld
.AND(m0_2
, retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
),
2269 brw_imm_ud(INTEL_MASK(16, 13)));
2271 /* Shift it up to bits 27:24. */
2272 fwa_bld
.SHL(m0_2
, m0_2
, brw_imm_ud(11));
2274 /* Set the Barrier Count and the enable bit */
2275 fwa_bld
.OR(m0_2
, m0_2
,
2276 brw_imm_ud(tcs_prog_data
->instances
<< 8 | (1 << 15)));
2278 bld
.emit(SHADER_OPCODE_BARRIER
, bld
.null_reg_ud(), m0
);
2282 case nir_intrinsic_load_input
:
2283 unreachable("nir_lower_io should never give us these.");
2286 case nir_intrinsic_load_per_vertex_input
: {
2287 fs_reg indirect_offset
= get_indirect_offset(instr
);
2288 unsigned imm_offset
= instr
->const_index
[0];
2290 const nir_src
&vertex_src
= instr
->src
[0];
2291 nir_const_value
*vertex_const
= nir_src_as_const_value(vertex_src
);
2298 /* Emit a MOV to resolve <0,1,0> regioning. */
2299 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2301 retype(brw_vec1_grf(1 + (vertex_const
->i32
[0] >> 3),
2302 vertex_const
->i32
[0] & 7),
2303 BRW_REGISTER_TYPE_UD
));
2304 } else if (tcs_prog_data
->instances
== 1 &&
2305 vertex_src
.is_ssa
&&
2306 vertex_src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
&&
2307 nir_instr_as_intrinsic(vertex_src
.ssa
->parent_instr
)->intrinsic
== nir_intrinsic_load_invocation_id
) {
2308 /* For the common case of only 1 instance, an array index of
2309 * gl_InvocationID means reading g1. Skip all the indirect work.
2311 icp_handle
= retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
);
2313 /* The vertex index is non-constant. We need to use indirect
2314 * addressing to fetch the proper URB handle.
2316 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2318 /* Each ICP handle is a single DWord (4 bytes) */
2319 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2320 bld
.SHL(vertex_offset_bytes
,
2321 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2324 /* Start at g1. We might read up to 4 registers. */
2325 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2326 fs_reg(brw_vec8_grf(1, 0)), vertex_offset_bytes
,
2327 brw_imm_ud(4 * REG_SIZE
));
2330 if (indirect_offset
.file
== BAD_FILE
) {
2331 /* Constant indexing - use global offset. */
2332 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2333 inst
->offset
= imm_offset
;
2335 inst
->base_mrf
= -1;
2336 inst
->regs_written
= instr
->num_components
;
2338 /* Indirect indexing - use per-slot offsets as well. */
2339 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2340 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2341 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2343 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2344 inst
->offset
= imm_offset
;
2345 inst
->base_mrf
= -1;
2347 inst
->regs_written
= instr
->num_components
;
2350 /* Copy the temporary to the destination to deal with writemasking.
2352 * Also attempt to deal with gl_PointSize being in the .w component.
2354 if (inst
->offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
2355 inst
->dst
= bld
.vgrf(dst
.type
, 4);
2356 inst
->regs_written
= 4;
2357 bld
.MOV(dst
, offset(inst
->dst
, bld
, 3));
2362 case nir_intrinsic_load_output
:
2363 case nir_intrinsic_load_per_vertex_output
: {
2364 fs_reg indirect_offset
= get_indirect_offset(instr
);
2365 unsigned imm_offset
= instr
->const_index
[0];
2368 if (indirect_offset
.file
== BAD_FILE
) {
2369 /* Replicate the patch handle to all enabled channels */
2370 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2371 bld
.MOV(patch_handle
,
2372 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
));
2374 if (imm_offset
== 0) {
2375 /* This is a read of gl_TessLevelInner[], which lives in the
2376 * Patch URB header. The layout depends on the domain.
2378 dst
.type
= BRW_REGISTER_TYPE_F
;
2379 switch (tcs_key
->tes_primitive_mode
) {
2381 /* DWords 3-2 (reversed) */
2382 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
2384 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, patch_handle
);
2387 inst
->base_mrf
= -1;
2388 inst
->regs_written
= 4;
2390 /* dst.xy = tmp.wz */
2391 bld
.MOV(dst
, offset(tmp
, bld
, 3));
2392 bld
.MOV(offset(dst
, bld
, 1), offset(tmp
, bld
, 2));
2396 /* DWord 4; hardcode offset = 1 and regs_written = 1 */
2397 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, patch_handle
);
2400 inst
->base_mrf
= -1;
2401 inst
->regs_written
= 1;
2404 /* All channels are undefined. */
2407 unreachable("Bogus tessellation domain");
2409 } else if (imm_offset
== 1) {
2410 /* This is a read of gl_TessLevelOuter[], which lives in the
2411 * Patch URB header. The layout depends on the domain.
2413 dst
.type
= BRW_REGISTER_TYPE_F
;
2415 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
2416 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, patch_handle
);
2419 inst
->base_mrf
= -1;
2420 inst
->regs_written
= 4;
2422 /* Reswizzle: WZYX */
2424 offset(tmp
, bld
, 3),
2425 offset(tmp
, bld
, 2),
2426 offset(tmp
, bld
, 1),
2427 offset(tmp
, bld
, 0),
2430 unsigned num_components
;
2431 switch (tcs_key
->tes_primitive_mode
) {
2439 /* Isolines are not reversed; swizzle .zw -> .xy */
2440 srcs
[0] = offset(tmp
, bld
, 2);
2441 srcs
[1] = offset(tmp
, bld
, 3);
2445 unreachable("Bogus tessellation domain");
2447 bld
.LOAD_PAYLOAD(dst
, srcs
, num_components
, 0);
2449 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, patch_handle
);
2450 inst
->offset
= imm_offset
;
2452 inst
->base_mrf
= -1;
2453 inst
->regs_written
= instr
->num_components
;
2456 /* Indirect indexing - use per-slot offsets as well. */
2457 const fs_reg srcs
[] = {
2458 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2461 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2462 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2464 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2465 inst
->offset
= imm_offset
;
2467 inst
->base_mrf
= -1;
2468 inst
->regs_written
= instr
->num_components
;
2473 case nir_intrinsic_store_output
:
2474 case nir_intrinsic_store_per_vertex_output
: {
2475 fs_reg value
= get_nir_src(instr
->src
[0]);
2476 fs_reg indirect_offset
= get_indirect_offset(instr
);
2477 unsigned imm_offset
= instr
->const_index
[0];
2478 unsigned swiz
= BRW_SWIZZLE_XYZW
;
2479 unsigned mask
= instr
->const_index
[1];
2480 unsigned header_regs
= 0;
2482 srcs
[header_regs
++] = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
);
2484 if (indirect_offset
.file
!= BAD_FILE
) {
2485 srcs
[header_regs
++] = indirect_offset
;
2486 } else if (!is_passthrough_shader
) {
2487 if (imm_offset
== 0) {
2488 value
.type
= BRW_REGISTER_TYPE_F
;
2490 mask
&= (1 << tesslevel_inner_components(tcs_key
->tes_primitive_mode
)) - 1;
2492 /* This is a write to gl_TessLevelInner[], which lives in the
2493 * Patch URB header. The layout depends on the domain.
2495 switch (tcs_key
->tes_primitive_mode
) {
2497 /* gl_TessLevelInner[].xy lives at DWords 3-2 (reversed).
2498 * We use an XXYX swizzle to reverse put .xy in the .wz
2499 * channels, and use a .zw writemask.
2501 mask
= writemask_for_backwards_vector(mask
);
2502 swiz
= BRW_SWIZZLE4(0, 0, 1, 0);
2505 /* gl_TessLevelInner[].x lives at DWord 4, so we set the
2506 * writemask to X and bump the URB offset by 1.
2511 /* Skip; gl_TessLevelInner[] doesn't exist for isolines. */
2514 unreachable("Bogus tessellation domain");
2516 } else if (imm_offset
== 1) {
2517 /* This is a write to gl_TessLevelOuter[] which lives in the
2518 * Patch URB Header at DWords 4-7. However, it's reversed, so
2519 * instead of .xyzw we have .wzyx.
2521 value
.type
= BRW_REGISTER_TYPE_F
;
2523 mask
&= (1 << tesslevel_outer_components(tcs_key
->tes_primitive_mode
)) - 1;
2525 if (tcs_key
->tes_primitive_mode
== GL_ISOLINES
) {
2526 /* Isolines .xy should be stored in .zw, in order. */
2527 swiz
= BRW_SWIZZLE4(0, 0, 0, 1);
2530 /* Other domains are reversed; store .wzyx instead of .xyzw */
2531 swiz
= BRW_SWIZZLE_WZYX
;
2532 mask
= writemask_for_backwards_vector(mask
);
2540 unsigned num_components
= _mesa_fls(mask
);
2543 if (mask
!= WRITEMASK_XYZW
) {
2544 srcs
[header_regs
++] = brw_imm_ud(mask
<< 16);
2545 opcode
= indirect_offset
.file
!= BAD_FILE
?
2546 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2547 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2549 opcode
= indirect_offset
.file
!= BAD_FILE
?
2550 SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2551 SHADER_OPCODE_URB_WRITE_SIMD8
;
2554 for (unsigned i
= 0; i
< num_components
; i
++) {
2555 if (mask
& (1 << i
))
2556 srcs
[header_regs
+ i
] = offset(value
, bld
, BRW_GET_SWZ(swiz
, i
));
2559 unsigned mlen
= header_regs
+ num_components
;
2562 bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
2563 bld
.LOAD_PAYLOAD(payload
, srcs
, mlen
, header_regs
);
2565 fs_inst
*inst
= bld
.emit(opcode
, bld
.null_reg_ud(), payload
);
2566 inst
->offset
= imm_offset
;
2568 inst
->base_mrf
= -1;
2573 nir_emit_intrinsic(bld
, instr
);
2579 fs_visitor::nir_emit_tes_intrinsic(const fs_builder
&bld
,
2580 nir_intrinsic_instr
*instr
)
2582 assert(stage
== MESA_SHADER_TESS_EVAL
);
2583 struct brw_tes_prog_data
*tes_prog_data
= (struct brw_tes_prog_data
*) prog_data
;
2586 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2587 dest
= get_nir_dest(instr
->dest
);
2589 switch (instr
->intrinsic
) {
2590 case nir_intrinsic_load_primitive_id
:
2591 bld
.MOV(dest
, fs_reg(brw_vec1_grf(0, 1)));
2593 case nir_intrinsic_load_tess_coord
:
2594 /* gl_TessCoord is part of the payload in g1-3 */
2595 for (unsigned i
= 0; i
< 3; i
++) {
2596 bld
.MOV(offset(dest
, bld
, i
), fs_reg(brw_vec8_grf(1 + i
, 0)));
2600 case nir_intrinsic_load_tess_level_outer
:
2601 /* When the TES reads gl_TessLevelOuter, we ensure that the patch header
2602 * appears as a push-model input. So, we can simply use the ATTR file
2603 * rather than issuing URB read messages. The data is stored in the
2604 * high DWords in reverse order - DWord 7 contains .x, DWord 6 contains
2607 switch (tes_prog_data
->domain
) {
2608 case BRW_TESS_DOMAIN_QUAD
:
2609 for (unsigned i
= 0; i
< 4; i
++)
2610 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2612 case BRW_TESS_DOMAIN_TRI
:
2613 for (unsigned i
= 0; i
< 3; i
++)
2614 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2616 case BRW_TESS_DOMAIN_ISOLINE
:
2617 for (unsigned i
= 0; i
< 2; i
++)
2618 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2623 case nir_intrinsic_load_tess_level_inner
:
2624 /* When the TES reads gl_TessLevelInner, we ensure that the patch header
2625 * appears as a push-model input. So, we can simply use the ATTR file
2626 * rather than issuing URB read messages.
2628 switch (tes_prog_data
->domain
) {
2629 case BRW_TESS_DOMAIN_QUAD
:
2630 bld
.MOV(dest
, component(fs_reg(ATTR
, 0), 3));
2631 bld
.MOV(offset(dest
, bld
, 1), component(fs_reg(ATTR
, 0), 2));
2633 case BRW_TESS_DOMAIN_TRI
:
2634 bld
.MOV(dest
, component(fs_reg(ATTR
, 0), 4));
2636 case BRW_TESS_DOMAIN_ISOLINE
:
2637 /* ignore - value is undefined */
2642 case nir_intrinsic_load_input
:
2643 case nir_intrinsic_load_per_vertex_input
: {
2644 fs_reg indirect_offset
= get_indirect_offset(instr
);
2645 unsigned imm_offset
= instr
->const_index
[0];
2648 if (indirect_offset
.file
== BAD_FILE
) {
2649 /* Arbitrarily only push up to 32 vec4 slots worth of data,
2650 * which is 16 registers (since each holds 2 vec4 slots).
2652 const unsigned max_push_slots
= 32;
2653 if (imm_offset
< max_push_slots
) {
2654 fs_reg src
= fs_reg(ATTR
, imm_offset
/ 2, dest
.type
);
2655 for (int i
= 0; i
< instr
->num_components
; i
++) {
2656 bld
.MOV(offset(dest
, bld
, i
),
2657 component(src
, 4 * (imm_offset
% 2) + i
));
2659 tes_prog_data
->base
.urb_read_length
=
2660 MAX2(tes_prog_data
->base
.urb_read_length
,
2661 DIV_ROUND_UP(imm_offset
+ 1, 2));
2663 /* Replicate the patch handle to all enabled channels */
2664 const fs_reg srcs
[] = {
2665 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)
2667 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2668 bld
.LOAD_PAYLOAD(patch_handle
, srcs
, ARRAY_SIZE(srcs
), 0);
2670 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dest
, patch_handle
);
2672 inst
->offset
= imm_offset
;
2673 inst
->base_mrf
= -1;
2674 inst
->regs_written
= instr
->num_components
;
2677 /* Indirect indexing - use per-slot offsets as well. */
2678 const fs_reg srcs
[] = {
2679 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2682 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2683 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2685 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dest
, payload
);
2687 inst
->offset
= imm_offset
;
2688 inst
->base_mrf
= -1;
2689 inst
->regs_written
= instr
->num_components
;
2694 nir_emit_intrinsic(bld
, instr
);
2700 fs_visitor::nir_emit_gs_intrinsic(const fs_builder
&bld
,
2701 nir_intrinsic_instr
*instr
)
2703 assert(stage
== MESA_SHADER_GEOMETRY
);
2704 fs_reg indirect_offset
;
2707 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2708 dest
= get_nir_dest(instr
->dest
);
2710 switch (instr
->intrinsic
) {
2711 case nir_intrinsic_load_primitive_id
:
2712 assert(stage
== MESA_SHADER_GEOMETRY
);
2713 assert(((struct brw_gs_prog_data
*)prog_data
)->include_primitive_id
);
2714 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
2715 retype(fs_reg(brw_vec8_grf(2, 0)), BRW_REGISTER_TYPE_UD
));
2718 case nir_intrinsic_load_input
:
2719 unreachable("load_input intrinsics are invalid for the GS stage");
2721 case nir_intrinsic_load_per_vertex_input
:
2722 emit_gs_input_load(dest
, instr
->src
[0], instr
->const_index
[0],
2723 instr
->src
[1], instr
->num_components
);
2726 case nir_intrinsic_emit_vertex_with_counter
:
2727 emit_gs_vertex(instr
->src
[0], instr
->const_index
[0]);
2730 case nir_intrinsic_end_primitive_with_counter
:
2731 emit_gs_end_primitive(instr
->src
[0]);
2734 case nir_intrinsic_set_vertex_count
:
2735 bld
.MOV(this->final_gs_vertex_count
, get_nir_src(instr
->src
[0]));
2738 case nir_intrinsic_load_invocation_id
: {
2739 fs_reg val
= nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
2740 assert(val
.file
!= BAD_FILE
);
2741 dest
.type
= val
.type
;
2747 nir_emit_intrinsic(bld
, instr
);
2753 fs_visitor::nir_emit_fs_intrinsic(const fs_builder
&bld
,
2754 nir_intrinsic_instr
*instr
)
2756 assert(stage
== MESA_SHADER_FRAGMENT
);
2757 struct brw_wm_prog_data
*wm_prog_data
=
2758 (struct brw_wm_prog_data
*) prog_data
;
2759 const struct brw_wm_prog_key
*wm_key
= (const struct brw_wm_prog_key
*) key
;
2762 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2763 dest
= get_nir_dest(instr
->dest
);
2765 switch (instr
->intrinsic
) {
2766 case nir_intrinsic_load_front_face
:
2767 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
2768 *emit_frontfacing_interpolation());
2771 case nir_intrinsic_load_sample_pos
: {
2772 fs_reg sample_pos
= nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
2773 assert(sample_pos
.file
!= BAD_FILE
);
2774 dest
.type
= sample_pos
.type
;
2775 bld
.MOV(dest
, sample_pos
);
2776 bld
.MOV(offset(dest
, bld
, 1), offset(sample_pos
, bld
, 1));
2780 case nir_intrinsic_load_helper_invocation
:
2781 case nir_intrinsic_load_sample_mask_in
:
2782 case nir_intrinsic_load_sample_id
: {
2783 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
2784 fs_reg val
= nir_system_values
[sv
];
2785 assert(val
.file
!= BAD_FILE
);
2786 dest
.type
= val
.type
;
2791 case nir_intrinsic_discard
:
2792 case nir_intrinsic_discard_if
: {
2793 /* We track our discarded pixels in f0.1. By predicating on it, we can
2794 * update just the flag bits that aren't yet discarded. If there's no
2795 * condition, we emit a CMP of g0 != g0, so all currently executing
2796 * channels will get turned off.
2799 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
2800 cmp
= bld
.CMP(bld
.null_reg_f(), get_nir_src(instr
->src
[0]),
2801 brw_imm_d(0), BRW_CONDITIONAL_Z
);
2803 fs_reg some_reg
= fs_reg(retype(brw_vec8_grf(0, 0),
2804 BRW_REGISTER_TYPE_UW
));
2805 cmp
= bld
.CMP(bld
.null_reg_f(), some_reg
, some_reg
, BRW_CONDITIONAL_NZ
);
2807 cmp
->predicate
= BRW_PREDICATE_NORMAL
;
2808 cmp
->flag_subreg
= 1;
2810 if (devinfo
->gen
>= 6) {
2811 emit_discard_jump();
2816 case nir_intrinsic_interp_var_at_centroid
:
2817 case nir_intrinsic_interp_var_at_sample
:
2818 case nir_intrinsic_interp_var_at_offset
: {
2819 /* Handle ARB_gpu_shader5 interpolation intrinsics
2821 * It's worth a quick word of explanation as to why we handle the full
2822 * variable-based interpolation intrinsic rather than a lowered version
2823 * with like we do for other inputs. We have to do that because the way
2824 * we set up inputs doesn't allow us to use the already setup inputs for
2825 * interpolation. At the beginning of the shader, we go through all of
2826 * the input variables and do the initial interpolation and put it in
2827 * the nir_inputs array based on its location as determined in
2828 * nir_lower_io. If the input isn't used, dead code cleans up and
2829 * everything works fine. However, when we get to the ARB_gpu_shader5
2830 * interpolation intrinsics, we need to reinterpolate the input
2831 * differently. If we used an intrinsic that just had an index it would
2832 * only give us the offset into the nir_inputs array. However, this is
2833 * useless because that value is post-interpolation and we need
2834 * pre-interpolation. In order to get the actual location of the bits
2835 * we get from the vertex fetching hardware, we need the variable.
2837 wm_prog_data
->pulls_bary
= true;
2839 fs_reg dst_xy
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 2);
2840 const glsl_interp_qualifier interpolation
=
2841 (glsl_interp_qualifier
) instr
->variables
[0]->var
->data
.interpolation
;
2843 switch (instr
->intrinsic
) {
2844 case nir_intrinsic_interp_var_at_centroid
:
2845 emit_pixel_interpolater_send(bld
,
2846 FS_OPCODE_INTERPOLATE_AT_CENTROID
,
2853 case nir_intrinsic_interp_var_at_sample
: {
2854 if (!wm_key
->multisample_fbo
) {
2855 /* From the ARB_gpu_shader5 specification:
2856 * "If multisample buffers are not available, the input varying
2857 * will be evaluated at the center of the pixel."
2859 emit_pixel_interpolater_send(bld
,
2860 FS_OPCODE_INTERPOLATE_AT_CENTROID
,
2868 nir_const_value
*const_sample
= nir_src_as_const_value(instr
->src
[0]);
2871 unsigned msg_data
= const_sample
->i32
[0] << 4;
2873 emit_pixel_interpolater_send(bld
,
2874 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
2877 brw_imm_ud(msg_data
),
2880 const fs_reg sample_src
= retype(get_nir_src(instr
->src
[0]),
2881 BRW_REGISTER_TYPE_UD
);
2883 if (nir_src_is_dynamically_uniform(instr
->src
[0])) {
2884 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
2885 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
2886 bld
.exec_all().group(1, 0)
2887 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
2888 emit_pixel_interpolater_send(bld
,
2889 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
2895 /* Make a loop that sends a message to the pixel interpolater
2896 * for the sample number in each live channel. If there are
2897 * multiple channels with the same sample number then these
2898 * will be handled simultaneously with a single interation of
2901 bld
.emit(BRW_OPCODE_DO
);
2903 /* Get the next live sample number into sample_id_reg */
2904 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
2906 /* Set the flag register so that we can perform the send
2907 * message on all channels that have the same sample number
2909 bld
.CMP(bld
.null_reg_ud(),
2910 sample_src
, sample_id
,
2911 BRW_CONDITIONAL_EQ
);
2912 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
2913 bld
.exec_all().group(1, 0)
2914 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
2916 emit_pixel_interpolater_send(bld
,
2917 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
2922 set_predicate(BRW_PREDICATE_NORMAL
, inst
);
2924 /* Continue the loop if there are any live channels left */
2925 set_predicate_inv(BRW_PREDICATE_NORMAL
,
2927 bld
.emit(BRW_OPCODE_WHILE
));
2934 case nir_intrinsic_interp_var_at_offset
: {
2935 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
2937 const bool flip
= !wm_key
->render_to_fbo
;
2940 unsigned off_x
= MIN2((int)(const_offset
->f32
[0] * 16), 7) & 0xf;
2941 unsigned off_y
= MIN2((int)(const_offset
->f32
[1] * 16 *
2942 (flip
? -1 : 1)), 7) & 0xf;
2944 emit_pixel_interpolater_send(bld
,
2945 FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
,
2948 brw_imm_ud(off_x
| (off_y
<< 4)),
2951 fs_reg src
= vgrf(glsl_type::ivec2_type
);
2952 fs_reg offset_src
= retype(get_nir_src(instr
->src
[0]),
2953 BRW_REGISTER_TYPE_F
);
2954 for (int i
= 0; i
< 2; i
++) {
2955 fs_reg temp
= vgrf(glsl_type::float_type
);
2956 bld
.MUL(temp
, offset(offset_src
, bld
, i
), brw_imm_f(16.0f
));
2957 fs_reg itemp
= vgrf(glsl_type::int_type
);
2959 bld
.MOV(itemp
, (i
== 1 && flip
) ? negate(temp
) : temp
);
2961 /* Clamp the upper end of the range to +7/16.
2962 * ARB_gpu_shader5 requires that we support a maximum offset
2963 * of +0.5, which isn't representable in a S0.4 value -- if
2964 * we didn't clamp it, we'd end up with -8/16, which is the
2965 * opposite of what the shader author wanted.
2967 * This is legal due to ARB_gpu_shader5's quantization
2970 * "Not all values of <offset> may be supported; x and y
2971 * offsets may be rounded to fixed-point values with the
2972 * number of fraction bits given by the
2973 * implementation-dependent constant
2974 * FRAGMENT_INTERPOLATION_OFFSET_BITS"
2976 set_condmod(BRW_CONDITIONAL_L
,
2977 bld
.SEL(offset(src
, bld
, i
), itemp
, brw_imm_d(7)));
2980 const enum opcode opcode
= FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
;
2981 emit_pixel_interpolater_send(bld
,
2992 unreachable("Invalid intrinsic");
2995 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
2996 fs_reg src
= interp_reg(instr
->variables
[0]->var
->data
.location
, j
);
2997 src
.type
= dest
.type
;
2999 bld
.emit(FS_OPCODE_LINTERP
, dest
, dst_xy
, src
);
3000 dest
= offset(dest
, bld
, 1);
3005 nir_emit_intrinsic(bld
, instr
);
3011 fs_visitor::nir_emit_cs_intrinsic(const fs_builder
&bld
,
3012 nir_intrinsic_instr
*instr
)
3014 assert(stage
== MESA_SHADER_COMPUTE
);
3015 struct brw_cs_prog_data
*cs_prog_data
=
3016 (struct brw_cs_prog_data
*) prog_data
;
3019 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3020 dest
= get_nir_dest(instr
->dest
);
3022 switch (instr
->intrinsic
) {
3023 case nir_intrinsic_barrier
:
3025 cs_prog_data
->uses_barrier
= true;
3028 case nir_intrinsic_load_local_invocation_id
:
3029 case nir_intrinsic_load_work_group_id
: {
3030 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
3031 fs_reg val
= nir_system_values
[sv
];
3032 assert(val
.file
!= BAD_FILE
);
3033 dest
.type
= val
.type
;
3034 for (unsigned i
= 0; i
< 3; i
++)
3035 bld
.MOV(offset(dest
, bld
, i
), offset(val
, bld
, i
));
3039 case nir_intrinsic_load_num_work_groups
: {
3040 const unsigned surface
=
3041 cs_prog_data
->binding_table
.work_groups_start
;
3043 cs_prog_data
->uses_num_work_groups
= true;
3045 fs_reg surf_index
= brw_imm_ud(surface
);
3046 brw_mark_surface_used(prog_data
, surface
);
3048 /* Read the 3 GLuint components of gl_NumWorkGroups */
3049 for (unsigned i
= 0; i
< 3; i
++) {
3050 fs_reg read_result
=
3051 emit_untyped_read(bld
, surf_index
,
3053 1 /* dims */, 1 /* size */,
3054 BRW_PREDICATE_NONE
);
3055 read_result
.type
= dest
.type
;
3056 bld
.MOV(dest
, read_result
);
3057 dest
= offset(dest
, bld
, 1);
3062 case nir_intrinsic_shared_atomic_add
:
3063 nir_emit_shared_atomic(bld
, BRW_AOP_ADD
, instr
);
3065 case nir_intrinsic_shared_atomic_imin
:
3066 nir_emit_shared_atomic(bld
, BRW_AOP_IMIN
, instr
);
3068 case nir_intrinsic_shared_atomic_umin
:
3069 nir_emit_shared_atomic(bld
, BRW_AOP_UMIN
, instr
);
3071 case nir_intrinsic_shared_atomic_imax
:
3072 nir_emit_shared_atomic(bld
, BRW_AOP_IMAX
, instr
);
3074 case nir_intrinsic_shared_atomic_umax
:
3075 nir_emit_shared_atomic(bld
, BRW_AOP_UMAX
, instr
);
3077 case nir_intrinsic_shared_atomic_and
:
3078 nir_emit_shared_atomic(bld
, BRW_AOP_AND
, instr
);
3080 case nir_intrinsic_shared_atomic_or
:
3081 nir_emit_shared_atomic(bld
, BRW_AOP_OR
, instr
);
3083 case nir_intrinsic_shared_atomic_xor
:
3084 nir_emit_shared_atomic(bld
, BRW_AOP_XOR
, instr
);
3086 case nir_intrinsic_shared_atomic_exchange
:
3087 nir_emit_shared_atomic(bld
, BRW_AOP_MOV
, instr
);
3089 case nir_intrinsic_shared_atomic_comp_swap
:
3090 nir_emit_shared_atomic(bld
, BRW_AOP_CMPWR
, instr
);
3093 case nir_intrinsic_load_shared
: {
3094 assert(devinfo
->gen
>= 7);
3096 fs_reg surf_index
= brw_imm_ud(GEN7_BTI_SLM
);
3098 /* Get the offset to read from */
3100 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3102 offset_reg
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0]);
3104 offset_reg
= vgrf(glsl_type::uint_type
);
3106 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
3107 brw_imm_ud(instr
->const_index
[0]));
3110 /* Read the vector */
3111 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, offset_reg
,
3113 instr
->num_components
,
3114 BRW_PREDICATE_NONE
);
3115 read_result
.type
= dest
.type
;
3116 for (int i
= 0; i
< instr
->num_components
; i
++)
3117 bld
.MOV(offset(dest
, bld
, i
), offset(read_result
, bld
, i
));
3122 case nir_intrinsic_store_shared
: {
3123 assert(devinfo
->gen
>= 7);
3126 fs_reg surf_index
= brw_imm_ud(GEN7_BTI_SLM
);
3129 fs_reg val_reg
= get_nir_src(instr
->src
[0]);
3132 unsigned writemask
= instr
->const_index
[1];
3134 /* Combine groups of consecutive enabled channels in one write
3135 * message. We use ffs to find the first enabled channel and then ffs on
3136 * the bit-inverse, down-shifted writemask to determine the length of
3137 * the block of enabled bits.
3140 unsigned first_component
= ffs(writemask
) - 1;
3141 unsigned length
= ffs(~(writemask
>> first_component
)) - 1;
3144 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3146 offset_reg
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0] +
3147 4 * first_component
);
3149 offset_reg
= vgrf(glsl_type::uint_type
);
3151 retype(get_nir_src(instr
->src
[1]), BRW_REGISTER_TYPE_UD
),
3152 brw_imm_ud(instr
->const_index
[0] + 4 * first_component
));
3155 emit_untyped_write(bld
, surf_index
, offset_reg
,
3156 offset(val_reg
, bld
, first_component
),
3157 1 /* dims */, length
,
3158 BRW_PREDICATE_NONE
);
3160 /* Clear the bits in the writemask that we just wrote, then try
3161 * again to see if more channels are left.
3163 writemask
&= (15 << (first_component
+ length
));
3170 nir_emit_intrinsic(bld
, instr
);
3176 fs_visitor::nir_emit_intrinsic(const fs_builder
&bld
, nir_intrinsic_instr
*instr
)
3179 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3180 dest
= get_nir_dest(instr
->dest
);
3182 switch (instr
->intrinsic
) {
3183 case nir_intrinsic_atomic_counter_inc
:
3184 case nir_intrinsic_atomic_counter_dec
:
3185 case nir_intrinsic_atomic_counter_read
: {
3186 /* Get the arguments of the atomic intrinsic. */
3187 const fs_reg offset
= get_nir_src(instr
->src
[0]);
3188 const unsigned surface
= (stage_prog_data
->binding_table
.abo_start
+
3189 instr
->const_index
[0]);
3192 /* Emit a surface read or atomic op. */
3193 switch (instr
->intrinsic
) {
3194 case nir_intrinsic_atomic_counter_read
:
3195 tmp
= emit_untyped_read(bld
, brw_imm_ud(surface
), offset
, 1, 1);
3198 case nir_intrinsic_atomic_counter_inc
:
3199 tmp
= emit_untyped_atomic(bld
, brw_imm_ud(surface
), offset
, fs_reg(),
3200 fs_reg(), 1, 1, BRW_AOP_INC
);
3203 case nir_intrinsic_atomic_counter_dec
:
3204 tmp
= emit_untyped_atomic(bld
, brw_imm_ud(surface
), offset
, fs_reg(),
3205 fs_reg(), 1, 1, BRW_AOP_PREDEC
);
3209 unreachable("Unreachable");
3212 /* Assign the result. */
3213 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
), tmp
);
3215 /* Mark the surface as used. */
3216 brw_mark_surface_used(stage_prog_data
, surface
);
3220 case nir_intrinsic_image_load
:
3221 case nir_intrinsic_image_store
:
3222 case nir_intrinsic_image_atomic_add
:
3223 case nir_intrinsic_image_atomic_min
:
3224 case nir_intrinsic_image_atomic_max
:
3225 case nir_intrinsic_image_atomic_and
:
3226 case nir_intrinsic_image_atomic_or
:
3227 case nir_intrinsic_image_atomic_xor
:
3228 case nir_intrinsic_image_atomic_exchange
:
3229 case nir_intrinsic_image_atomic_comp_swap
: {
3230 using namespace image_access
;
3232 /* Get the referenced image variable and type. */
3233 const nir_variable
*var
= instr
->variables
[0]->var
;
3234 const glsl_type
*type
= var
->type
->without_array();
3235 const brw_reg_type base_type
= get_image_base_type(type
);
3237 /* Get some metadata from the image intrinsic. */
3238 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
3239 const unsigned arr_dims
= type
->sampler_array
? 1 : 0;
3240 const unsigned surf_dims
= type
->coordinate_components() - arr_dims
;
3241 const unsigned format
= var
->data
.image
.format
;
3243 /* Get the arguments of the image intrinsic. */
3244 const fs_reg image
= get_nir_image_deref(instr
->variables
[0]);
3245 const fs_reg addr
= retype(get_nir_src(instr
->src
[0]),
3246 BRW_REGISTER_TYPE_UD
);
3247 const fs_reg src0
= (info
->num_srcs
>= 3 ?
3248 retype(get_nir_src(instr
->src
[2]), base_type
) :
3250 const fs_reg src1
= (info
->num_srcs
>= 4 ?
3251 retype(get_nir_src(instr
->src
[3]), base_type
) :
3255 /* Emit an image load, store or atomic op. */
3256 if (instr
->intrinsic
== nir_intrinsic_image_load
)
3257 tmp
= emit_image_load(bld
, image
, addr
, surf_dims
, arr_dims
, format
);
3259 else if (instr
->intrinsic
== nir_intrinsic_image_store
)
3260 emit_image_store(bld
, image
, addr
, src0
, surf_dims
, arr_dims
,
3261 var
->data
.image
.write_only
? GL_NONE
: format
);
3264 tmp
= emit_image_atomic(bld
, image
, addr
, src0
, src1
,
3265 surf_dims
, arr_dims
, info
->dest_components
,
3266 get_image_atomic_op(instr
->intrinsic
, type
));
3268 /* Assign the result. */
3269 for (unsigned c
= 0; c
< info
->dest_components
; ++c
)
3270 bld
.MOV(offset(retype(dest
, base_type
), bld
, c
),
3271 offset(tmp
, bld
, c
));
3275 case nir_intrinsic_memory_barrier_atomic_counter
:
3276 case nir_intrinsic_memory_barrier_buffer
:
3277 case nir_intrinsic_memory_barrier_image
:
3278 case nir_intrinsic_memory_barrier
: {
3279 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 16 / dispatch_width
);
3280 bld
.emit(SHADER_OPCODE_MEMORY_FENCE
, tmp
)
3285 case nir_intrinsic_group_memory_barrier
:
3286 case nir_intrinsic_memory_barrier_shared
:
3287 /* We treat these workgroup-level barriers as no-ops. This should be
3288 * safe at present and as long as:
3290 * - Memory access instructions are not subsequently reordered by the
3291 * compiler back-end.
3293 * - All threads from a given compute shader workgroup fit within a
3294 * single subslice and therefore talk to the same HDC shared unit
3295 * what supposedly guarantees ordering and coherency between threads
3296 * from the same workgroup. This may change in the future when we
3297 * start splitting workgroups across multiple subslices.
3299 * - The context is not in fault-and-stream mode, which could cause
3300 * memory transactions (including to SLM) prior to the barrier to be
3301 * replayed after the barrier if a pagefault occurs. This shouldn't
3302 * be a problem up to and including SKL because fault-and-stream is
3303 * not usable due to hardware issues, but that's likely to change in
3308 case nir_intrinsic_shader_clock
: {
3309 /* We cannot do anything if there is an event, so ignore it for now */
3310 fs_reg shader_clock
= get_timestamp(bld
);
3311 const fs_reg srcs
[] = { shader_clock
.set_smear(0), shader_clock
.set_smear(1) };
3313 bld
.LOAD_PAYLOAD(dest
, srcs
, ARRAY_SIZE(srcs
), 0);
3317 case nir_intrinsic_image_size
: {
3318 /* Get the referenced image variable and type. */
3319 const nir_variable
*var
= instr
->variables
[0]->var
;
3320 const glsl_type
*type
= var
->type
->without_array();
3322 /* Get the size of the image. */
3323 const fs_reg image
= get_nir_image_deref(instr
->variables
[0]);
3324 const fs_reg size
= offset(image
, bld
, BRW_IMAGE_PARAM_SIZE_OFFSET
);
3326 /* For 1DArray image types, the array index is stored in the Z component.
3327 * Fix this by swizzling the Z component to the Y component.
3329 const bool is_1d_array_image
=
3330 type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_1D
&&
3331 type
->sampler_array
;
3333 /* For CubeArray images, we should count the number of cubes instead
3334 * of the number of faces. Fix it by dividing the (Z component) by 6.
3336 const bool is_cube_array_image
=
3337 type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
3338 type
->sampler_array
;
3340 /* Copy all the components. */
3341 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
3342 for (unsigned c
= 0; c
< info
->dest_components
; ++c
) {
3343 if ((int)c
>= type
->coordinate_components()) {
3344 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3346 } else if (c
== 1 && is_1d_array_image
) {
3347 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3348 offset(size
, bld
, 2));
3349 } else if (c
== 2 && is_cube_array_image
) {
3350 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
,
3351 offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3352 offset(size
, bld
, c
), brw_imm_d(6));
3354 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3355 offset(size
, bld
, c
));
3362 case nir_intrinsic_image_samples
:
3363 /* The driver does not support multi-sampled images. */
3364 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), brw_imm_d(1));
3367 case nir_intrinsic_load_uniform
: {
3368 /* Offsets are in bytes but they should always be multiples of 4 */
3369 assert(instr
->const_index
[0] % 4 == 0);
3371 fs_reg
src(UNIFORM
, instr
->const_index
[0] / 4, dest
.type
);
3373 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3375 /* Offsets are in bytes but they should always be multiples of 4 */
3376 assert(const_offset
->u32
[0] % 4 == 0);
3377 src
.reg_offset
= const_offset
->u32
[0] / 4;
3379 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3380 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
3383 fs_reg indirect
= retype(get_nir_src(instr
->src
[0]),
3384 BRW_REGISTER_TYPE_UD
);
3386 /* We need to pass a size to the MOV_INDIRECT but we don't want it to
3387 * go past the end of the uniform. In order to keep the n'th
3388 * component from running past, we subtract off the size of all but
3389 * one component of the vector.
3391 assert(instr
->const_index
[1] >=
3392 instr
->num_components
* (int) type_sz(dest
.type
));
3393 unsigned read_size
= instr
->const_index
[1] -
3394 (instr
->num_components
- 1) * type_sz(dest
.type
);
3396 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3397 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
3398 offset(dest
, bld
, j
), offset(src
, bld
, j
),
3399 indirect
, brw_imm_ud(read_size
));
3405 case nir_intrinsic_load_ubo
: {
3406 nir_const_value
*const_index
= nir_src_as_const_value(instr
->src
[0]);
3410 const unsigned index
= stage_prog_data
->binding_table
.ubo_start
+
3411 const_index
->u32
[0];
3412 surf_index
= brw_imm_ud(index
);
3413 brw_mark_surface_used(prog_data
, index
);
3415 /* The block index is not a constant. Evaluate the index expression
3416 * per-channel and add the base UBO index; we have to select a value
3417 * from any live channel.
3419 surf_index
= vgrf(glsl_type::uint_type
);
3420 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
3421 brw_imm_ud(stage_prog_data
->binding_table
.ubo_start
));
3422 surf_index
= bld
.emit_uniformize(surf_index
);
3424 /* Assume this may touch any UBO. It would be nice to provide
3425 * a tighter bound, but the array information is already lowered away.
3427 brw_mark_surface_used(prog_data
,
3428 stage_prog_data
->binding_table
.ubo_start
+
3429 nir
->info
.num_ubos
- 1);
3432 /* Number of 32-bit slots in the type */
3433 unsigned type_slots
= MAX2(1, type_sz(dest
.type
) / 4);
3435 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3436 if (const_offset
== NULL
) {
3437 fs_reg base_offset
= retype(get_nir_src(instr
->src
[1]),
3438 BRW_REGISTER_TYPE_UD
);
3440 for (int i
= 0; i
< instr
->num_components
; i
++)
3441 VARYING_PULL_CONSTANT_LOAD(bld
, offset(dest
, bld
, i
), surf_index
,
3442 base_offset
, i
* type_sz(dest
.type
));
3444 /* Even if we are loading doubles, a pull constant load will load
3445 * a 32-bit vec4, so should only reserve vgrf space for that. If we
3446 * need to load a full dvec4 we will have to emit 2 loads. This is
3447 * similar to demote_pull_constants(), except that in that case we
3448 * see individual accesses to each component of the vector and then
3449 * we let CSE deal with duplicate loads. Here we see a vector access
3450 * and we have to split it if necessary.
3452 fs_reg packed_consts
= vgrf(glsl_type::float_type
);
3453 packed_consts
.type
= dest
.type
;
3455 unsigned const_offset_aligned
= const_offset
->u32
[0] & ~15;
3457 /* A vec4 only contains half of a dvec4, if we need more than 2
3458 * components of a dvec4 we will have to issue another load for
3459 * components z and w.
3462 if (type_slots
== 1)
3463 num_components
= instr
->num_components
;
3465 num_components
= MIN2(2, instr
->num_components
);
3467 /* The computation of num_components doesn't take into account
3468 * misalignment, which should be okay according to std140 vector
3471 assert(const_offset
->u32
[0] % 16 +
3472 type_sz(dest
.type
) * num_components
<= 16);
3474 int remaining_components
= instr
->num_components
;
3475 while (remaining_components
> 0) {
3476 /* Read the vec4 from a 16-byte aligned offset */
3477 struct brw_reg const_offset_reg
= brw_imm_ud(const_offset_aligned
);
3478 bld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
3479 retype(packed_consts
, BRW_REGISTER_TYPE_F
),
3480 surf_index
, const_offset_reg
);
3482 const fs_reg consts
= byte_offset(packed_consts
, (const_offset
->u32
[0] % 16));
3483 unsigned dest_offset
= instr
->num_components
- remaining_components
;
3485 /* XXX: This doesn't update the sub-16B offset across iterations of
3486 * the loop, which should work for std140 vector alignment rules.
3488 assert(dest_offset
== 0 || const_offset
->u32
[0] % 16 == 0);
3490 for (int i
= 0; i
< num_components
; i
++)
3491 bld
.MOV(offset(dest
, bld
, i
+ dest_offset
), component(consts
, i
));
3493 /* If this is a large enough 64-bit load, we will need to emit
3496 remaining_components
-= num_components
;
3497 assert(remaining_components
== 0 ||
3498 (remaining_components
<= 2 && type_slots
== 2));
3499 num_components
= remaining_components
;
3500 const_offset_aligned
+= 16;
3506 case nir_intrinsic_load_ssbo
: {
3507 assert(devinfo
->gen
>= 7);
3509 nir_const_value
*const_uniform_block
=
3510 nir_src_as_const_value(instr
->src
[0]);
3513 if (const_uniform_block
) {
3514 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
3515 const_uniform_block
->u32
[0];
3516 surf_index
= brw_imm_ud(index
);
3517 brw_mark_surface_used(prog_data
, index
);
3519 surf_index
= vgrf(glsl_type::uint_type
);
3520 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
3521 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3523 /* Assume this may touch any UBO. It would be nice to provide
3524 * a tighter bound, but the array information is already lowered away.
3526 brw_mark_surface_used(prog_data
,
3527 stage_prog_data
->binding_table
.ssbo_start
+
3528 nir
->info
.num_ssbos
- 1);
3532 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3534 offset_reg
= brw_imm_ud(const_offset
->u32
[0]);
3536 offset_reg
= get_nir_src(instr
->src
[1]);
3539 /* Read the vector */
3540 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, offset_reg
,
3542 instr
->num_components
,
3543 BRW_PREDICATE_NONE
);
3544 read_result
.type
= dest
.type
;
3545 for (int i
= 0; i
< instr
->num_components
; i
++)
3546 bld
.MOV(offset(dest
, bld
, i
), offset(read_result
, bld
, i
));
3551 case nir_intrinsic_load_input
: {
3553 if (stage
== MESA_SHADER_VERTEX
) {
3554 src
= fs_reg(ATTR
, instr
->const_index
[0], dest
.type
);
3556 src
= offset(retype(nir_inputs
, dest
.type
), bld
,
3557 instr
->const_index
[0]);
3560 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3561 assert(const_offset
&& "Indirect input loads not allowed");
3562 src
= offset(src
, bld
, const_offset
->u32
[0]);
3564 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3565 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
3570 case nir_intrinsic_store_ssbo
: {
3571 assert(devinfo
->gen
>= 7);
3575 nir_const_value
*const_uniform_block
=
3576 nir_src_as_const_value(instr
->src
[1]);
3577 if (const_uniform_block
) {
3578 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
3579 const_uniform_block
->u32
[0];
3580 surf_index
= brw_imm_ud(index
);
3581 brw_mark_surface_used(prog_data
, index
);
3583 surf_index
= vgrf(glsl_type::uint_type
);
3584 bld
.ADD(surf_index
, get_nir_src(instr
->src
[1]),
3585 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3587 brw_mark_surface_used(prog_data
,
3588 stage_prog_data
->binding_table
.ssbo_start
+
3589 nir
->info
.num_ssbos
- 1);
3593 fs_reg val_reg
= get_nir_src(instr
->src
[0]);
3596 unsigned writemask
= instr
->const_index
[0];
3598 /* Combine groups of consecutive enabled channels in one write
3599 * message. We use ffs to find the first enabled channel and then ffs on
3600 * the bit-inverse, down-shifted writemask to determine the length of
3601 * the block of enabled bits.
3604 unsigned first_component
= ffs(writemask
) - 1;
3605 unsigned length
= ffs(~(writemask
>> first_component
)) - 1;
3608 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[2]);
3610 offset_reg
= brw_imm_ud(const_offset
->u32
[0] + 4 * first_component
);
3612 offset_reg
= vgrf(glsl_type::uint_type
);
3614 retype(get_nir_src(instr
->src
[2]), BRW_REGISTER_TYPE_UD
),
3615 brw_imm_ud(4 * first_component
));
3618 emit_untyped_write(bld
, surf_index
, offset_reg
,
3619 offset(val_reg
, bld
, first_component
),
3620 1 /* dims */, length
,
3621 BRW_PREDICATE_NONE
);
3623 /* Clear the bits in the writemask that we just wrote, then try
3624 * again to see if more channels are left.
3626 writemask
&= (15 << (first_component
+ length
));
3631 case nir_intrinsic_store_output
: {
3632 fs_reg src
= get_nir_src(instr
->src
[0]);
3633 fs_reg new_dest
= offset(retype(nir_outputs
, src
.type
), bld
,
3634 instr
->const_index
[0]);
3636 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3637 assert(const_offset
&& "Indirect output stores not allowed");
3638 new_dest
= offset(new_dest
, bld
, const_offset
->u32
[0]);
3640 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3641 bld
.MOV(offset(new_dest
, bld
, j
), offset(src
, bld
, j
));
3646 case nir_intrinsic_ssbo_atomic_add
:
3647 nir_emit_ssbo_atomic(bld
, BRW_AOP_ADD
, instr
);
3649 case nir_intrinsic_ssbo_atomic_imin
:
3650 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMIN
, instr
);
3652 case nir_intrinsic_ssbo_atomic_umin
:
3653 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMIN
, instr
);
3655 case nir_intrinsic_ssbo_atomic_imax
:
3656 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMAX
, instr
);
3658 case nir_intrinsic_ssbo_atomic_umax
:
3659 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMAX
, instr
);
3661 case nir_intrinsic_ssbo_atomic_and
:
3662 nir_emit_ssbo_atomic(bld
, BRW_AOP_AND
, instr
);
3664 case nir_intrinsic_ssbo_atomic_or
:
3665 nir_emit_ssbo_atomic(bld
, BRW_AOP_OR
, instr
);
3667 case nir_intrinsic_ssbo_atomic_xor
:
3668 nir_emit_ssbo_atomic(bld
, BRW_AOP_XOR
, instr
);
3670 case nir_intrinsic_ssbo_atomic_exchange
:
3671 nir_emit_ssbo_atomic(bld
, BRW_AOP_MOV
, instr
);
3673 case nir_intrinsic_ssbo_atomic_comp_swap
:
3674 nir_emit_ssbo_atomic(bld
, BRW_AOP_CMPWR
, instr
);
3677 case nir_intrinsic_get_buffer_size
: {
3678 nir_const_value
*const_uniform_block
= nir_src_as_const_value(instr
->src
[0]);
3679 unsigned ssbo_index
= const_uniform_block
? const_uniform_block
->u32
[0] : 0;
3680 int reg_width
= dispatch_width
/ 8;
3683 fs_reg source
= brw_imm_d(0);
3685 int mlen
= 1 * reg_width
;
3687 /* A resinfo's sampler message is used to get the buffer size.
3688 * The SIMD8's writeback message consists of four registers and
3689 * SIMD16's writeback message consists of 8 destination registers
3690 * (two per each component), although we are only interested on the
3691 * first component, where resinfo returns the buffer size for
3694 int regs_written
= 4 * mlen
;
3695 fs_reg src_payload
= fs_reg(VGRF
, alloc
.allocate(mlen
),
3696 BRW_REGISTER_TYPE_UD
);
3697 bld
.LOAD_PAYLOAD(src_payload
, &source
, 1, 0);
3698 fs_reg buffer_size
= fs_reg(VGRF
, alloc
.allocate(regs_written
),
3699 BRW_REGISTER_TYPE_UD
);
3700 const unsigned index
= prog_data
->binding_table
.ssbo_start
+ ssbo_index
;
3701 fs_inst
*inst
= bld
.emit(FS_OPCODE_GET_BUFFER_SIZE
, buffer_size
,
3702 src_payload
, brw_imm_ud(index
));
3703 inst
->header_size
= 0;
3705 inst
->regs_written
= regs_written
;
3707 bld
.MOV(retype(dest
, buffer_size
.type
), buffer_size
);
3709 brw_mark_surface_used(prog_data
, index
);
3714 unreachable("unknown intrinsic");
3719 fs_visitor::nir_emit_ssbo_atomic(const fs_builder
&bld
,
3720 int op
, nir_intrinsic_instr
*instr
)
3723 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3724 dest
= get_nir_dest(instr
->dest
);
3727 nir_const_value
*const_surface
= nir_src_as_const_value(instr
->src
[0]);
3728 if (const_surface
) {
3729 unsigned surf_index
= stage_prog_data
->binding_table
.ssbo_start
+
3730 const_surface
->u32
[0];
3731 surface
= brw_imm_ud(surf_index
);
3732 brw_mark_surface_used(prog_data
, surf_index
);
3734 surface
= vgrf(glsl_type::uint_type
);
3735 bld
.ADD(surface
, get_nir_src(instr
->src
[0]),
3736 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3738 /* Assume this may touch any SSBO. This is the same we do for other
3739 * UBO/SSBO accesses with non-constant surface.
3741 brw_mark_surface_used(prog_data
,
3742 stage_prog_data
->binding_table
.ssbo_start
+
3743 nir
->info
.num_ssbos
- 1);
3746 fs_reg offset
= get_nir_src(instr
->src
[1]);
3747 fs_reg data1
= get_nir_src(instr
->src
[2]);
3749 if (op
== BRW_AOP_CMPWR
)
3750 data2
= get_nir_src(instr
->src
[3]);
3752 /* Emit the actual atomic operation operation */
3754 fs_reg atomic_result
= emit_untyped_atomic(bld
, surface
, offset
,
3756 1 /* dims */, 1 /* rsize */,
3758 BRW_PREDICATE_NONE
);
3759 dest
.type
= atomic_result
.type
;
3760 bld
.MOV(dest
, atomic_result
);
3764 fs_visitor::nir_emit_shared_atomic(const fs_builder
&bld
,
3765 int op
, nir_intrinsic_instr
*instr
)
3768 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3769 dest
= get_nir_dest(instr
->dest
);
3771 fs_reg surface
= brw_imm_ud(GEN7_BTI_SLM
);
3772 fs_reg offset
= get_nir_src(instr
->src
[0]);
3773 fs_reg data1
= get_nir_src(instr
->src
[1]);
3775 if (op
== BRW_AOP_CMPWR
)
3776 data2
= get_nir_src(instr
->src
[2]);
3778 /* Emit the actual atomic operation operation */
3780 fs_reg atomic_result
= emit_untyped_atomic(bld
, surface
, offset
,
3782 1 /* dims */, 1 /* rsize */,
3784 BRW_PREDICATE_NONE
);
3785 dest
.type
= atomic_result
.type
;
3786 bld
.MOV(dest
, atomic_result
);
3790 fs_visitor::nir_emit_texture(const fs_builder
&bld
, nir_tex_instr
*instr
)
3792 unsigned texture
= instr
->texture_index
;
3793 unsigned sampler
= instr
->sampler_index
;
3795 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
3797 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture
);
3798 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_ud(sampler
);
3800 int lod_components
= 0;
3802 /* The hardware requires a LOD for buffer textures */
3803 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
3804 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_d(0);
3806 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
3807 fs_reg src
= get_nir_src(instr
->src
[i
].src
);
3808 switch (instr
->src
[i
].src_type
) {
3809 case nir_tex_src_bias
:
3810 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
3812 case nir_tex_src_comparitor
:
3813 srcs
[TEX_LOGICAL_SRC_SHADOW_C
] = retype(src
, BRW_REGISTER_TYPE_F
);
3815 case nir_tex_src_coord
:
3816 switch (instr
->op
) {
3818 case nir_texop_txf_ms
:
3819 case nir_texop_txf_ms_mcs
:
3820 case nir_texop_samples_identical
:
3821 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_D
);
3824 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_F
);
3828 case nir_tex_src_ddx
:
3829 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
3830 lod_components
= nir_tex_instr_src_size(instr
, i
);
3832 case nir_tex_src_ddy
:
3833 srcs
[TEX_LOGICAL_SRC_LOD2
] = retype(src
, BRW_REGISTER_TYPE_F
);
3835 case nir_tex_src_lod
:
3836 switch (instr
->op
) {
3838 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_UD
);
3841 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_D
);
3844 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
3848 case nir_tex_src_ms_index
:
3849 srcs
[TEX_LOGICAL_SRC_SAMPLE_INDEX
] = retype(src
, BRW_REGISTER_TYPE_UD
);
3852 case nir_tex_src_offset
: {
3853 nir_const_value
*const_offset
=
3854 nir_src_as_const_value(instr
->src
[i
].src
);
3856 unsigned header_bits
= brw_texture_offset(const_offset
->i32
, 3);
3857 if (header_bits
!= 0)
3858 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
] = brw_imm_ud(header_bits
);
3860 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
] =
3861 retype(src
, BRW_REGISTER_TYPE_D
);
3866 case nir_tex_src_projector
:
3867 unreachable("should be lowered");
3869 case nir_tex_src_texture_offset
: {
3870 /* Figure out the highest possible texture index and mark it as used */
3871 uint32_t max_used
= texture
+ instr
->texture_array_size
- 1;
3872 if (instr
->op
== nir_texop_tg4
&& devinfo
->gen
< 8) {
3873 max_used
+= stage_prog_data
->binding_table
.gather_texture_start
;
3875 max_used
+= stage_prog_data
->binding_table
.texture_start
;
3877 brw_mark_surface_used(prog_data
, max_used
);
3879 /* Emit code to evaluate the actual indexing expression */
3880 fs_reg tmp
= vgrf(glsl_type::uint_type
);
3881 bld
.ADD(tmp
, src
, brw_imm_ud(texture
));
3882 srcs
[TEX_LOGICAL_SRC_SURFACE
] = bld
.emit_uniformize(tmp
);
3886 case nir_tex_src_sampler_offset
: {
3887 /* Emit code to evaluate the actual indexing expression */
3888 fs_reg tmp
= vgrf(glsl_type::uint_type
);
3889 bld
.ADD(tmp
, src
, brw_imm_ud(sampler
));
3890 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = bld
.emit_uniformize(tmp
);
3894 case nir_tex_src_ms_mcs
:
3895 assert(instr
->op
== nir_texop_txf_ms
);
3896 srcs
[TEX_LOGICAL_SRC_MCS
] = retype(src
, BRW_REGISTER_TYPE_D
);
3900 unreachable("unknown texture source");
3904 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BAD_FILE
&&
3905 (instr
->op
== nir_texop_txf_ms
||
3906 instr
->op
== nir_texop_samples_identical
)) {
3907 if (devinfo
->gen
>= 7 &&
3908 key_tex
->compressed_multisample_layout_mask
& (1 << texture
)) {
3909 srcs
[TEX_LOGICAL_SRC_MCS
] =
3910 emit_mcs_fetch(srcs
[TEX_LOGICAL_SRC_COORDINATE
],
3911 instr
->coord_components
,
3912 srcs
[TEX_LOGICAL_SRC_SURFACE
]);
3914 srcs
[TEX_LOGICAL_SRC_MCS
] = brw_imm_ud(0u);
3918 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_d(instr
->coord_components
);
3919 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_d(lod_components
);
3921 if (instr
->op
== nir_texop_query_levels
) {
3922 /* textureQueryLevels() is implemented in terms of TXS so we need to
3923 * pass a valid LOD argument.
3925 assert(srcs
[TEX_LOGICAL_SRC_LOD
].file
== BAD_FILE
);
3926 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_ud(0u);
3930 switch (instr
->op
) {
3932 opcode
= SHADER_OPCODE_TEX_LOGICAL
;
3935 opcode
= FS_OPCODE_TXB_LOGICAL
;
3938 opcode
= SHADER_OPCODE_TXL_LOGICAL
;
3941 opcode
= SHADER_OPCODE_TXD_LOGICAL
;
3944 opcode
= SHADER_OPCODE_TXF_LOGICAL
;
3946 case nir_texop_txf_ms
:
3947 if ((key_tex
->msaa_16
& (1 << sampler
)))
3948 opcode
= SHADER_OPCODE_TXF_CMS_W_LOGICAL
;
3950 opcode
= SHADER_OPCODE_TXF_CMS_LOGICAL
;
3952 case nir_texop_txf_ms_mcs
:
3953 opcode
= SHADER_OPCODE_TXF_MCS_LOGICAL
;
3955 case nir_texop_query_levels
:
3957 opcode
= SHADER_OPCODE_TXS_LOGICAL
;
3960 opcode
= SHADER_OPCODE_LOD_LOGICAL
;
3963 if (srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
!= BAD_FILE
&&
3964 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
!= IMM
)
3965 opcode
= SHADER_OPCODE_TG4_OFFSET_LOGICAL
;
3967 opcode
= SHADER_OPCODE_TG4_LOGICAL
;
3969 case nir_texop_texture_samples
: {
3970 fs_reg dst
= retype(get_nir_dest(instr
->dest
), BRW_REGISTER_TYPE_D
);
3972 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
, 4);
3973 fs_inst
*inst
= bld
.emit(SHADER_OPCODE_SAMPLEINFO
, tmp
,
3974 bld
.vgrf(BRW_REGISTER_TYPE_D
, 1),
3975 srcs
[TEX_LOGICAL_SRC_SURFACE
],
3976 srcs
[TEX_LOGICAL_SRC_SURFACE
]);
3978 inst
->header_size
= 1;
3979 inst
->base_mrf
= -1;
3980 inst
->regs_written
= 4 * (dispatch_width
/ 8);
3982 /* Pick off the one component we care about */
3986 case nir_texop_samples_identical
: {
3987 fs_reg dst
= retype(get_nir_dest(instr
->dest
), BRW_REGISTER_TYPE_D
);
3989 /* If mcs is an immediate value, it means there is no MCS. In that case
3990 * just return false.
3992 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BRW_IMMEDIATE_VALUE
) {
3993 bld
.MOV(dst
, brw_imm_ud(0u));
3994 } else if ((key_tex
->msaa_16
& (1 << sampler
))) {
3995 fs_reg tmp
= vgrf(glsl_type::uint_type
);
3996 bld
.OR(tmp
, srcs
[TEX_LOGICAL_SRC_MCS
],
3997 offset(srcs
[TEX_LOGICAL_SRC_MCS
], bld
, 1));
3998 bld
.CMP(dst
, tmp
, brw_imm_ud(0u), BRW_CONDITIONAL_EQ
);
4000 bld
.CMP(dst
, srcs
[TEX_LOGICAL_SRC_MCS
], brw_imm_ud(0u),
4001 BRW_CONDITIONAL_EQ
);
4006 unreachable("unknown texture opcode");
4009 fs_reg dst
= bld
.vgrf(brw_type_for_nir_type(instr
->dest_type
), 4);
4010 fs_inst
*inst
= bld
.emit(opcode
, dst
, srcs
, ARRAY_SIZE(srcs
));
4012 const unsigned dest_size
= nir_tex_instr_dest_size(instr
);
4013 if (devinfo
->gen
>= 9 &&
4014 instr
->op
!= nir_texop_tg4
&& instr
->op
!= nir_texop_query_levels
) {
4015 unsigned write_mask
= instr
->dest
.is_ssa
?
4016 nir_ssa_def_components_read(&instr
->dest
.ssa
):
4017 (1 << dest_size
) - 1;
4018 assert(write_mask
!= 0); /* dead code should have been eliminated */
4019 inst
->regs_written
= _mesa_fls(write_mask
) * dispatch_width
/ 8;
4021 inst
->regs_written
= 4 * dispatch_width
/ 8;
4024 if (srcs
[TEX_LOGICAL_SRC_SHADOW_C
].file
!= BAD_FILE
)
4025 inst
->shadow_compare
= true;
4027 if (srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
== IMM
)
4028 inst
->offset
= srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].ud
;
4030 if (instr
->op
== nir_texop_tg4
) {
4031 if (instr
->component
== 1 &&
4032 key_tex
->gather_channel_quirk_mask
& (1 << texture
)) {
4033 /* gather4 sampler is broken for green channel on RG32F --
4034 * we must ask for blue instead.
4036 inst
->offset
|= 2 << 16;
4038 inst
->offset
|= instr
->component
<< 16;
4041 if (devinfo
->gen
== 6)
4042 emit_gen6_gather_wa(key_tex
->gen6_gather_wa
[texture
], dst
);
4046 for (unsigned i
= 0; i
< dest_size
; i
++)
4047 nir_dest
[i
] = offset(dst
, bld
, i
);
4049 bool is_cube_array
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
4052 if (instr
->op
== nir_texop_query_levels
) {
4053 /* # levels is in .w */
4054 nir_dest
[0] = offset(dst
, bld
, 3);
4055 } else if (instr
->op
== nir_texop_txs
&& dest_size
>= 3 &&
4056 (devinfo
->gen
< 7 || is_cube_array
)) {
4057 fs_reg depth
= offset(dst
, bld
, 2);
4058 fs_reg fixed_depth
= vgrf(glsl_type::int_type
);
4060 if (is_cube_array
) {
4061 /* fixup #layers for cube map arrays */
4062 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, fixed_depth
, depth
, brw_imm_d(6));
4063 } else if (devinfo
->gen
< 7) {
4064 /* Gen4-6 return 0 instead of 1 for single layer surfaces. */
4065 bld
.emit_minmax(fixed_depth
, depth
, brw_imm_d(1), BRW_CONDITIONAL_GE
);
4068 nir_dest
[2] = fixed_depth
;
4071 bld
.LOAD_PAYLOAD(get_nir_dest(instr
->dest
), nir_dest
, dest_size
, 0);
4075 fs_visitor::nir_emit_jump(const fs_builder
&bld
, nir_jump_instr
*instr
)
4077 switch (instr
->type
) {
4078 case nir_jump_break
:
4079 bld
.emit(BRW_OPCODE_BREAK
);
4081 case nir_jump_continue
:
4082 bld
.emit(BRW_OPCODE_CONTINUE
);
4084 case nir_jump_return
:
4086 unreachable("unknown jump");
4091 * This helper takes the result of a load operation that reads 32-bit elements
4099 * and shuffles the data to get this:
4106 * Which is exactly what we want if the load is reading 64-bit components
4107 * like doubles, where x represents the low 32-bit of the x double component
4108 * and y represents the high 32-bit of the x double component (likewise with
4109 * z and w for double component y). The parameter @components represents
4110 * the number of 64-bit components present in @src. This would typically be
4111 * 2 at most, since we can only fit 2 double elements in the result of a
4114 * Notice that @dst and @src can be the same register.
4117 shuffle_32bit_load_result_to_64bit_data(const fs_builder
&bld
,
4120 uint32_t components
)
4122 assert(type_sz(src
.type
) == 4);
4123 assert(type_sz(dst
.type
) == 8);
4125 /* A temporary that we will use to shuffle the 32-bit data of each
4126 * component in the vector into valid 64-bit data. We can't write directly
4127 * to dst because dst can be (and would usually be) the same as src
4128 * and in that case the first MOV in the loop below would overwrite the
4129 * data read in the second MOV.
4131 fs_reg tmp
= bld
.vgrf(dst
.type
);
4133 for (unsigned i
= 0; i
< components
; i
++) {
4134 const fs_reg component_i
= offset(src
, bld
, 2 * i
);
4136 bld
.MOV(subscript(tmp
, src
.type
, 0), component_i
);
4137 bld
.MOV(subscript(tmp
, src
.type
, 1), offset(component_i
, bld
, 1));
4139 bld
.MOV(offset(dst
, bld
, i
), tmp
);